In radio astronomy, signals from radio telescopes are transformed into images to observe celestial objects, or sources. However, these images, called dirty images, contain real sources as well as artifacts due to signal sparsity and other factors. Therefore, radio interferometric image reconstruction is performed on dirty images, aiming to produce clean images in which artifacts are reduced and real sources are recovered. So far, existing methods have limited success on recovering faint sources, preserving detailed structures, and eliminating artifacts. In this paper, we present VIC-DDPM, a Visibility and Image Conditioned Denoising Diffusion Probabilistic Model. Our main idea is to use both the original visibility data in the spectral domain and dirty images in the spatial domain to guide the image generation process with DDPM. This way, we can leverage DDPM to generate fine details and eliminate noise, while utilizing visibility data to separate signals from noise and retaining spatial information in dirty images. We have conducted experiments in comparison with both traditional methods and recent deep learning based approaches. Our results show that our method significantly improves the resulting images by reducing artifacts, preserving fine details, and recovering dim sources. This advancement further facilitates radio astronomical data analysis tasks on celestial phenomena. Our code is available at https://github.com/RapidsAtHKUST/VIC-DDPM.

Training the machine learning (ML) models require a large amount of data; however, the capacity of these models is limited. Recent literature focuses on combining ML models' predictions with human experts. This setting is known as the human-in-the-loop or human-AI team setting. Human experts can complement the ML models as they are well-equipped with vast real-world experience and sometimes have access to private information that may not be accessible while training the ML model. Existing approaches for combining an expert and ML model either require end-to-end training of the combined model or require expert annotations for every task. Such end-to-end training requires a custom loss function and human annotations, which is cumbersome, results in slower convergence, and may adversely impact the ML model's accuracy. On the other hand, using expert annotations for every task is also cost-ineffective. We propose a novel technique that optimizes the cost of seeking the expert's advice while utilizing the ML model's predictions to improve accuracy. Our model considers two intrinsic parameters: the expert's cost for each prediction and the misclassification cost of the combined human-AI model. Further, we present the impact of group-wise calibration on the combined model that improves the overall model's performance. Experimental results on our combined model with group-wise calibration show a significant increase in accuracy with limited expert advice against different established ML models for the image classification task. In addition, the combined model's accuracy is always greater than the ML model, irrespective of the expert's accuracy, the expert's cost, and the misclassification cost.

We present a new supervised learning technique for the Variational AutoEncoder (VAE) that allows it to learn a causally disentangled representation and generate causally disentangled outcomes simultaneously. We call this approach Causally Disentangled Generation (CDG). CDG is a generative model that accurately decodes an output based on a causally disentangled representation. Our research demonstrates that adding supervised regularization to the encoder alone is insufficient for achieving a generative model with CDG, even for a simple task. Therefore, we explore the necessary and sufficient conditions for achieving CDG within a specific model. Additionally, we introduce a universal metric for evaluating the causal disentanglement of a generative model. Empirical results from both image and tabular datasets support our findings.

This paper introduces a motivated agent scheme that enables an agent to create its own goals using prior knowledge about its environment. A motivated agent operates in a dynamically changing environment and is capable of setting and achieving its own goals, as well as those set by the designer. The agent has access to additional knowledge about the environment, which is represented in associative semantic memory. This memory is constructed based on ANAKG associative knowledge graphs, which have been shown to have several advantages over other semantic memories for processing symbolic sequential inputs. They are easy to organize and train. In this paper, we demonstrate that a motivated agent with semantic memory learns to achieve its goals more easily and utilizes environmental resources more effectively. To simplify comparisons with reinforcement learning, we represent the environment as an environmental graph that shows the principles governing it. By exploring the environment, the agent learns these principles and can use them to accomplish its tasks. Our experiments and tests confirm our claims about the higher efficiency of agents with memory. Moreover, an extensive comparison with reinforcement learning agents highlights the advantages of motivated learning over reinforcement learning.

Unsupervised Video Anomaly Detection (UVAD) utilizes completely unlabeled videos for training without any human intervention. Due to the existence of unlabeled abnormal videos in the training data, the performance of UVAD has a large gap compared with semi-supervised VAD, which only uses normal videos for training. To address the problem of insufficient ability of the existing UVAD methods to learn normality and reduce the negative impact of abnormal events, this paper proposes a novel Enhanced Spatio-temporal Self-selective Learning (ESSL) framework for UVAD. This framework is designed for capturing both the appearance and motion features through effective network structures by solving the spatial and temporal jigsaw puzzles. Specially, we develop a Self-selective Learning Module (SLM) for UVAD, which prevents the model learning abnormal features and enhances the model by selecting normal features. Experimental results on three benchmark datasets show that the proposed method not only surpasses the state-of-the-art UVAD works, but also achieves the performance comparable to the classic semi-supervised methods for video anomaly detection that needs normal videos selected manually.

Plan verification is the task of deciding whether a given plan is a solution to a planning problem. In this paper, we study the plan verification problem in the context of Hierarchical Task Network (HTN) planning, which has been proved to be NP-complete when partial order (PO) is involved. We will develop a novel SAT-based approach exploiting the data structures solution order graphs and path decomposition trees which encodes an HTN plan verification problem as a SAT one. We show in our experiments that this new approach outperforms the current state-of-the-art (SOTA) planning-based approach for verifying plans for POHTN problems.

In various applications, ads are displayed together with prices, so as to provide a direct comparison among similar products or services. The price-displaying feature not only influences the consumers' decision, but also affects the bidding behavior of advertisers. In this paper, we study ad auctions with display prices from the perspective of mechanism design, in which advertisers are asked to submit both the product costs and the display prices of their commodities. We first provide a characterization for all individually rational and incentive-compatible mechanisms in the presence of display prices, then use it to design ad auctions in two scenarios. In the former scenario, the display prices are assumed to be exogenously determined. For this scenario, we derive the welfare-maximizing and revenue-maximizing auctions for any given display price profile. In the latter, advertisers are allowed to strategize their display prices freely. We investigate two families of allocation policies within the scenario and identify the equilibrium display prices accordingly. Our findings demonstrate the impact of display prices on the design of ad auctions, and highlight how platforms can utilize display price information to optimize the performance of ad delivery. 

Object detection is an essential part of many perception algorithms used in modern robotics applications. Unfortunately, the existing models share a tendency to assign high confidence scores for out-of-distribution (OOD) samples. Although OOD detection has been extensively studied in recent years among the computer vision community, most proposed solutions apply only to the image recognition task. Real-world applications such as perception in autonomous vehicle struggle with far more complex challenges than classification. In our work, we focus on the popular field of object detection, adapting Neuron Activation Patterns (NAP) for out-of-distribution samples detection in object detection (NAPTRON). Our experiments show that NAPTRON outperforms state-of-the-art methods, without the need to affect in-distribution (ID) performance. We evaluate the methods on two different OOD scenarios and three types of object detectors which makes the comparison the largest open-source benchmark for OOD object detection.

Cereal grain plays a crucial role in the human diet as a major source of essential nutrients. Grain Appearance Inspection (GAI) serves as an essential process to determine grain quality and facilitate grain circulation, storage and processing. However, GAI is routinely performed manually by inspectors with cumbersome and tedious procedures, which poses a significant bottleneck in grain safety and smart agriculture. Therefore, it is imperative to establish a fair and efficient GAI system to ensure grain safety, thus contributing to the achievement of ``Zero Hunger'' and ``Good Health and Well-being'' for ending poverty and other deprivations. In this paper, we endeavor to develop an automated GAI system:  AI4GrainInsp. By analyzing the distinctive characteristics of grain kernels, we formulate GAI as a ubiquitous problem: Anomaly Detection (AD), in which healthy and edible kernels are considered normal samples while damaged grains or unknown objects are regarded as anomalies. We further propose an AD model, called AD-GAI, which is trained using only normal samples yet can identify anomalies during inference. Moreover, we customize a prototype device for data acquisition and create a large-scale dataset including 220K high-quality images of wheat and maize kernels. Through extensive experiments, AD-GAI achieves considerable performance in comparison with advanced AD methods, and AI4GrainInsp has highly consistent performance compared to human experts and excels at inspection efficiency over 20x speedup. The dataset, code and models will be released at  (*******anonymous).

Image geolocalization is receiving increasing attention due to its importance in several applications, such as image retrieval, criminal investigations and fact-checking. Previous works focused on several instances of image geolocalization including place recognition, GPS coordinates estimation and country recognition. In this paper, we tackle an even more challenging problem, which is recognizing the city where an image has been taken. Due to the vast number of cities in the world, we cast the problem as a verification problem, whereby the system has to decide whether a certain image has been taken in a given city or not. In particular, we present a system that given a query image and a small set of images taken in a target city, decides if the query image has been shot in the target city or not. To allow the system to handle the case of images, taken in cities that have not been used during training, we use a Siamese network based on Vision Transformer as a backbone. The experiments we run prove the validity of the proposed system which outperforms solutions based on state-of-the-art techniques, even in the challenging case of images shot in different cities of the same country.

The goal of inductive logic programming is to induce a logic program (a set of logical rules) that generalises training examples. Inducing programs with many rules and literals is a major challenge. To tackle this challenge, we introduce an approach where we learn small 'non-separable' programs and combine them. We implement our approach in a generate, test, combine, and constrain loop. Our approach can learn optimal and recursive programs and perform predicate invention. Our experiments on multiple domains, including game playing and program synthesis, show that our approach can drastically outperform existing approaches in terms of predictive accuracies and learning times, sometimes reducing learning times from over one hour to a few seconds.

In this work, we present an unsupervised dual constraint contrastive method for efficiently fine-tuning the vision-language pre-trained (VLP) models that have achieved great success on various cross-modal tasks, since full fine-tune these pre-trained models is computationally expensive and tend to result in catastrophic forgetting restricted by the size and quality of labeled datasets. Our approach freezes the pre-trained VLP models as the fundamental, generalized, and transferable multimodal representation and incorporates lightweight parameters to learn domain and task-specific features without labeled data. We demonstrated that our unsupervised dual contrastive model performs better than previous fine-tuning methods on MS COCO and Flickr 30K datasets on the cross-modal retrieval task, with an even more pronounced improvement in zero-shot performance. Furthermore, experiments on the MOTIF dataset prove that our unsupervised approach remains effective when trained on a small, out-of-domain dataset without overfitting. As a plug-and-play method, our proposed method is agnostic to the underlying models and can be easily integrated with different VLP models, allowing for the potential incorporation of future advancements in VLP models.

Aspect-based sentiment classification is a crucial problem in fine-grained sentiment analysis, which aims to predict the sentiment polarity of the given aspect according to its context. Previous works have made remarkable progress in leveraging attention mechanism to extract opinion words for different aspects. However, a persistent challenge is the effective management of semantic mismatches, which stem from attention mechanisms that fall short in adequately aligning opinions words with their corresponding aspect in multi-aspect sentences. To address this issue, we propose a novel Aspect-oriented Opinion Alignment Network (AOAN) to capture the contextual association between opinion words and the corresponding aspect. Specifically, we first introduce a neighboring span enhanced module which highlights various compositions of neighboring words and given aspects. In addition, we design a multi-perspective attention mechanism that align relevant opinion information with respect to the given aspect. Extensive experiments on three benchmark datasets demonstrate that our model achieves state-of-the-art results. The source code is available at https://github.com/AONE-NLP/ABSA-AOAN.

Visual question answering (VQA) is a crucial yet challenging task in multimodal understanding. To correctly answer questions about an image, VQA models are required to comprehend the fine-grained semantics of both the image and the question. Recent advances have shown that both grid and region features contribute to improving the VQA performance, while grid features surprisingly outperform region features. However, grid features will inevitably induce visual semantic noise due to fine granularity. Besides, the ignorance of geometric relationships makes VQA models difficult to understand the object relative positions in the image and answer questions accurately. In this paper, we propose a visual enhancement network for VQA that leverages region features and position information to enhance grid features, thus generating richer visual grid semantics. First, the grid enhancement multi-head guided attention module utilizes regions around the grid to provide visual context, forming rich visual grid semantics and effectively compensating for the fine granularity of the grid. Second, a novel geometric perception multi-head self-attention is introduced to process two types of features, incorporating geometric relations such as relative direction between objects while exploring internal semantic interactions. Extensive experiments demonstrate that the proposed method can obtain competitive results over other strong baselines.

Cost partitioning is the foundation of today's strongest heuristics for optimal classical planning. However, computing a cost partitioning for each evaluated state is prohibitively expensive in practice. Thus, existing approaches make an approximation and compute a cost partitioning only for a set of sampled states, and then reuse the resulting heuristics for all other states evaluated during the search. In this paper, we present exact methods for cost partitioning heuristics based on linear programming that fully preserve heuristic accuracy while minimizing computational cost. Specifically, we focus on saturated post-hoc optimization and establish several sufficient conditions for when reusing a cost partitioning computed for one state is optimal for other states, mainly based on a sensitivity analysis of the underlying linear program formulation. Our experiments demonstrate that our theoretical results transfer into practice and that our exact cost partitioning algorithms are competitive with the strongest approximations currently available while needing fewer linear program evaluations.

Zero-shot Learning is the process of transferring knowledge (Cross-modal mapping relationship) of the seen classes to unseen classes. However, the realization of this knowledge transfer process relies on a large number of expensive labels. Even if the model has labels, Zero-shot learning also has the phenomenon of 'negative causality' in the process of Cross-modal alignment. In this paper, We propose an unsupervised learning model and leverage a counterfactual causal inference framework for cross-modal mapping relationship adjustment (CMRA). Specifically, we aim to regard images as cause and Wikipedia text as effect form a causal relationship diagram. First, it uses multiple attributes attention to learn images semantic attributes and corresponding Wikipedia text description word to form Cross-modal alignment. Then, using contrastive learning and stop-gradient are combined to form a new Cross-modal mapping relationship. Finally, we compare whether multiple attributes attention has consistency in the distribution of semantic attributes learned before and after image transformation and the attentions with larger distributions are eliminated using an inactivation strategy. This model evaluates the classification accuracy in AWA, CUB, APY, SUN. The experimental results show that the algorithm outperforms the state-of-the-art algorithms technology approaches.

Bayesian neural networks (BNNs) have drawn extensive interest because of their distinctive probabilistic representation framework. However, despite its recent success, little work focuses on the information-theoretic understanding of Bayesian neural networks. In this paper, we propose Information Bound as a metric of the amount of information in Bayesian neural networks. Different from mutual information on deterministic neural networks where modification of network structure or specific input data is usually necessary, Information Bound can be easily estimated on current Bayesian neural networks without any modification of network structures or training processes. By observing the trend of Information Bound during training, we demonstrate the existence of the ``critical period'' in Bayesian neural networks. Besides, we show that the Information Bound can be used to judge the confidence of the model prediction and to detect out-of-distribution datasets. Based on these observations of model interpretation, we propose Information Bound regularization and Information Bound variance regularization methods. The Information Bound regularization encourages models to learn the minimum necessary information and improves the model generality and robustness. The Information Bound variance regularization encourages models to learn more about complex samples with low Information Bound. Extensive experiments on KMNIST, Fashion-MNIST, CIFAR-10, and CIFAR-100 verify the effectiveness of the proposed regularization methods.

Transfer learning in Reinforcement Learning (RL) has been widely studied to overcome training challenges in Deep-RL, i.e., exploration cost, data availability and convergence time, by bootstrapping external knowledge to enhance learning phase. While this overcomes the training issues on a novice agent, a good understanding of the task by the expert agent is required for such a transfer to be effective. As an alternative, in this paper we propose Expert-Free Online Transfer Learning (EF-OnTL), an algorithm that enables expert-free real-time dynamic transfer learning in multi-agent system. No dedicated expert agent exists, and transfer source agent and knowledge to be transferred are dynamically selected at each transfer step based on agents’ performance and level of uncertainty. To improve uncertainty estimation, we also propose State Action Reward Next-State Random Network Distillation (sars-RND), an extension of RND that estimates uncertainty from RL agent-environment interaction. We demonstrate EF-OnTL effectiveness against a no-transfer scenario and state-of-the-art advice-based baselines, with and without expert agents, in three benchmark tasks: Cart-Pole, a grid-based Multi-Team Predator-Prey (MT-PP) and Half Field Offense (HFO). Our results show that EF-OnTL achieves overall comparable performance to that of advice-based approaches, while not requiring expert agents, external input, nor threshold tuning. EF-OnTL outperforms no-transfer with an improvement related to the complexity of the task addressed.

The popularity of Graph Neural Networks (GNNs) has soared due to their proficiency in encapsulating intricate connections among graphs by aggregating information from (the immediate) node neighbors. Nonetheless, when dealing with graphs displaying a high degree of heterophily, the GNN performance declines as the faraway nodes that are pertinent to the task at hand are ignored during node aggregation. To contend with the issue of integrating long-range neighbors that are relevant into the GNN node aggregation mechanism, this paper introduces a model called the Overlay Graph Neural Networks (OGN).  OGN is inspired by P2P overlay networks, where the idea is to find neighbor peers (nodes) that, although not directly connected to a given node (a peer), are semantically similar and could favorably improve both query routing and query results. In the context of OGN, the network refers to the graph, and the message passing procedure executed by a GNN to aggregate node features represents the routing process. OGN networks are built by stacking one or more overlay layers, each taking as input the graph and a node feature matrix either available or derivable (e.g., by analyzing the graph’s structure). Each overlay layer combines base embeddings, learned by considering node features and short-range node neighbors, with overlay embeddings computed by projecting nodes with similar features close in an overlay space and then aggregating (overlay) neighbor nodes via a sliding window attention mechanism. Base and overlay embeddings are combined to capture nodes’ immediate and global context in a graph. We evaluate OGN in a node classification task using state-of-the-art benchmarks and show that the OGN model is competitive with the advantage of being easily portable to any existing GNN model.

We study the computational complexity of identifying a small agreeable subset of items. A subset of items is agreeable if every agent does not prefer its complement set. We study the setting where agents can give arbitrary utilities to the items, can only approve or disapprove items, or rank the items with Borda scores. We prove all variants NP-hard, and perform a parameterized analysis regarding the natural parameters number of agents, number of items, and the upper bound on the size of the agreeable subset in question.

Graph Lottery Ticket (GLT), a combination of core subgraph and sparse subnetwork, has been proposed to mitigate the computational cost of deep Graph Neural Networks (GNNs) on large input graphs while preserving original performance. However, the winning GLTs in exisiting studies are obtained by applying iterative magnitude-based pruning (IMP) without re-evaluating and re-considering the pruned information, which disregards the dynamic changes in the significance of edges/weights during graph/model structure pruning, and thus limits the appeal of the winning tickets. In this paper, we formulate a conjecture, i.e., existing overlooked valuable information in the pruned graph connections and model parameters which can be re-grouped into GLT to enhance the final performance. Specifically, we propose an adversarial complementary erasing (ACE) framework to explore the valuable information from the pruned components, thereby developing a more powerful GLT, referred to as the ACE-GLT. The main idea is to mine valuable information from pruned edges/weights after each round of IMP, and employ the ACE technique to refine the GLT processing. Finally, experimental results demonstrate that our ACE-GLT outperforms existing methods for searching GLT in diverse tasks. Our code will be made publicly available.

As more decisions in our daily life become automated, the need  to have machine learning algorithms that make fair decisions increases. In fair representation learning we are tasked with finding a suitable representation of the data in which a sensitive variable is censored. Recent work aims to learn fair representations through adversarial learning. With this paper, we build upon this work by introducing a novel algorithm, ALFR++, to learn adversarial fair representations independent of the downstream task that may be encountered (supervised or unsupervised). We introduce balanced error rate modulation, to stabilize the interaction between actor and adversary. We couple it with a recently introduced censoring technique called stacking, to learn strong censored representations within a restricted hypothesis space. Results show that our algorithm improves on previous work in both censoring and reconstruction, allowing us to create fairer data without knowing the task beforehand.

Directed evolution is a widely-used strategy of protein engineering to improve protein function via mimicking natural mutation and selection. Machine learning-assisted directed evolution(MLDE) approaches aim to learn a fitness predictor, thereby efficiently searching for optimal mutants within the vast combinatorial mutation space. Since annotating mutants is both costly and labor-intensive, how to efficiently sample and utilize informative protein mutants to train the predictor is a critical problem in MLDE. Previous MLDE works just simply utilized pre-trained protein language models (PPLMs) for sampling without tailoring to the specific target protein of interest, which has not fully exploited the potential of PPLMs. In this work, we propose a novel method, the Actively-Finetuned Protein model for Directed Evolution(AFP-DE) which leverages PPLMs to actively sample and fine-tune themselves, continuously improving the model’s sampling and overall performance through iterations, to achieve efficiently directed protein evolution. Extensive experiments have shown the effectiveness of our method in generating optimal mutants with minimal annotation effort, outperforming previous works even with fewer annotated mutants, making it budget-friendly for biological experiments. 

Hierarchical reinforcement learning is an increasingly demanded resource for learning to make sequential decisions towards long term goals with successful credit assignment and temporal abstraction. Feudal hierarchies are among the most deployed frameworks. However, there is lack of formalism over the hierarchical structure and of theoretical guarantees. We formalize the common two-level feudal hierarchy as two Markov decision processes, with the one on the high-level being dependent on the policy executed at the low-level. Despite the non-stationarity raised by the dependency, we show that each of the processes presents stable behavior. We then build on the first result to show that, regardless of the convergent learning algorithm used for the low-level, convergence of both prediction and control algorithms at the high-level is guaranteed with probability 1. Our results contribute with theoretical support for the use of feudal hierarchies in combination with standard reinforcement learning methods at each level.

With the growth of networks, promoting products through social networks has become an important problem. For auctions in social networks, items are needed to be sold to agents in a network, where each agent  can bid and also diffuse the sale information to her neighbors. Thus, the agents' social relations are  intervened with their bids in the auctions. In network auctions, the classical VCG  mechanism fails to retain key properties. In order to better understand network auctions, in this paper, we characterize network auctions for the single-unit setting with respect to IR, WBB, IC, efficiency, and other properties. For example, we present sufficient conditions for mechanisms to be efficient and (weakly) incentive compatible. With the help of these properties and new concepts such as rewards, participation rewards, and so on, we show how to design efficient mechanisms to satisfy IC as much as possible, and IC mechanisms to maximize the revenue. Our results  provide insights into understanding auctions in social networks.

User-item interaction data in recommender systems is a form of dyadic relation, reflecting user preferences for specific items. To generate accurate recommendations, it is crucial to learn representations for both users and items. Recent multimodal recommendation models achieve higher accuracy by incorporating multimodal features, such as images and text descriptions. However, our experimental findings reveal that current multimodality fusion methods employed in state-of-the-art models may adversely affect recommendation performance without compromising model architectures. Moreover, these models seldom investigate internal relations between item-item and user-user interactions. In light of these findings, we propose a model that enhances the dyadic relations by learning Dual RepresentAtions of both users and items via constructing homogeneous Graphs for multimOdal recommeNdation. We name our model as DRAGON. Specifically, DRAGON constructs user-user graphs based on commonly interacted items and item-item graphs derived from item multimodal features. Graph learning on both the user-item heterogeneous and homogeneous graphs is used to obtain dual representations of users and items. To capture information from each modality, DRAGON employs an effective fusion method, attentive concatenation. Extensive experiments on three public datasets and eight baselines show that DRAGON can outperform the strongest baseline by 21.41% on average.

Non-IID data presents a significant challenge for federated learning, and personalized Federated Learning(FL) is a natural solution to address this challenge. Recently, Graph Neural Network (GNN) has recently emerged to model the complex client relationship using a client graph to refine personalized models. However, this approach depends on an existing client relation graph on the server, making it impractical unless this prerequisite is satisfied. Furthermore, noisy and missing connections in the original graph structures can degrade personalization performance. In this work, we propose an unsupervised structure learning approach to improve personalized FL, where the server learns a dynamic client graph through self-supervision and generates structure-based client representations. These representations are then broadcasted to users, regulating local training using the learned knowledge as an inductive bias. Empirical studies on benchmark datasets demonstrate the significant effectiveness of our approach and the high quality of the client graphs.

Large Language Models (LLMs) such as ChatGPT have received enormous attention over the past year, and are now used by hundreds of millions of people every day. The rapid adoption of this technology naturally raises serious questions about the possible biases that such models might exhibit. In this work, we tested one of these models (GPT-3) on a range of cognitive effects, which are systematic patterns that are usually found in human cognition. We found that such models are indeed prone to a number of human cognitive effects. Specifically, we show that the \emph{priming}, \emph{distance}, \emph{SNARC}, and \emph{size congruity} effects were presented with GPT-3, while the \emph{anchoring} effect is absent. We describe our methodology, and specifically the way we converted real-world experiments to text-based experiments. Finally, we speculate on the possible reasons why GPT-3 exhibits these effects, and discuss the question of whether they are imitated or reinvented.

We present a novel approach to address the multi-agent sparse contextual linear bandit problem, in which the feature vectors have a high dimension $d$ whereas the reward function depends on only a limited set of features - precisely $s_0 \ll d$. Furthermore, the learning follows under information-sharing constraints. The proposed method employs Lasso regression for dimension reduction, allowing each agent to independently estimate an approximate set of main dimensions and share that information with others depending on the network's structure. The information is then aggregated through a specific process and shared with all agents. Each agent then resolves the problem with ridge regression focusing solely on the extracted dimensions. We represent algorithms for both a star-shaped network and a peer-to-peer network. The approaches effectively reduce communication costs while ensuring minimal cumulative regret per agent. Theoretically, we show that our proposed methods have a probabilistic regret bound of order $\mathcal{O}(s_0 \log d + s_0 \sqrt{T})$, where $T$ is the time horizon. To our best knowledge, it is the first algorithm that tackles row-wise distributed data in sparse linear bandits, achieving comparable performance compared to the state-of-the-art single and multi-agent methods. Besides, it is widely applicable to high-dimensional multi-agent problems where efficient feature extraction is critical for minimizing regret. To validate the effectiveness of our approach, we present experimental results on both synthetic and real-world datasets.

Gradient clipping is a commonly used technique to stabilize the training process of neural networks. A growing body of studies has shown that gradient clipping is a promising technique for dealing with the heavy-tailed behavior that emerged in stochastic optimization as well. While gradient clipping is significant, its theoretical guarantees are scarce. Most theoretical guarantees only provide an in-expectation analysis and only focus on optimization performance. In this paper, we provide high probability analysis in the non-convex setting and derive the optimization bound and the generalization bound simultaneously for popular stochastic optimization algorithms with gradient clipping, including stochastic gradient descent and its variants of momentum and adaptive stepsizes. With the gradient clipping, we study a heavy-tailed assumption that the gradients only have bounded $\alpha$-th moments for some $\alpha \in (1, 2]$, which is much weaker than the standard bounded second-moment assumption. Overall, our study provides a relatively complete picture for the theoretical guarantee of stochastic optimization algorithms with clipping.

Explaining artificial intelligence models can be utilized to launch targeted adversarial attacks on text classification algorithms. Understanding the reasoning behind the model's decisions makes it easier to prepare such samples. Most of the current text-based adversarial attacks rely on brute-force by using SHAP approach to identify the importance of tokens in the samples, we modify the crucial ones to prepare targeted attacks. We base our results on experiments using 5 datasets. Our results show that our approach outperforms TextBugger and TextFooler, achieving better results with 4 out of 5 datasets against TextBugger, and 3 out of 5 datasets against TextFooler, while minimizing perturbation introduced to the texts. In particular, we managed to outperform the efficacy of TextFooler by over 1900% and TextBugger by over 200% on the WikiPL dataset, additionally keeping high cosine similarity between the original text sample and the adversarial example. The evaluation of the results was additionally supported through a survey to assess their quality and ensure that the text perturbations did not change the intended class according to subjective, human classification.

The learnware paradigm aims to construct a market of numerous well-performing machine learning models, which enables users to leverage these models to accomplish specific tasks without having to build models from scratch. Each learnware in the market is a model associated with a specification, representing the model's utility and enabling it to be identified according to future users' requirements. In the learnware paradigm, due to the vast and ever-increasing number of models in the market, a significant challenge is to identify helpful learnwares efficiently for a specific user task without leaking data privacy. However, existing identification methods require examining the whole market, which is computationally unaffordable in a large market. In this paper, we propose a new framework for identifying helpful learnwares without examining the whole market. Specifically, using the Reduced Kernel Mean Embedding (RKME) specification, we derive a novel learnware scoring criterion for assessing the helpfulness of a learnware, based on which we design an anchor-based framework to identify helpful learnwares by examining only a small portion of learnwares in the market. Theoretical analyses are provided for both the criterion and the anchor-based method. Empirical studies on market containing thousands of learnwares from real-world datasets confirm the effectiveness of our proposed approach.

Large-scale vision-language models (LVLMs) pre-trained on massive image-text pairs have achieved remarkable success in visual representations. However, existing paradigms to transfer LVLMs to downstream tasks encounter two primary challenges. Firstly, the text features remain fixed after being calculated and cannot be adjusted according to image features, which decreases the model's adaptability. Secondly, the model’s output solely depends on the similarity between the text and image features, leading to excessive reliance on LVLMs. To address these two challenges, we introduce a novel two-branch model named the Instance-Wise Adaptive Tuning  and Caching ( ATC) . Specifically, one branch implements our proposed ConditionNet, which guides image features to form an adaptive textual cache that adjusts based on image features,achieving instance-wise inference and improving the model's adaptability. The other branch introduces the similarities between images and incorporates a learnable visual cache, designed to decouple new and previous knowledge, allowing the model to acquire new knowledge while preserving prior knowledge. The model's output is jointly determined by the two branches, thus overcoming the limitations of existing methods that rely solely on LVLMs. Additionally, our method requires  limited computing resources to tune parameters, yet outperforms existing methods on 11 benchmark datasets.

In this study, we delve into the “short circuit” phenomenon observed in multiple-choice natural language reasoning tasks, where models tend to make accurate choices without properly considering the context of the question. To better understand this phenomenon, we propose white-box and black-box proxy tests as investigative tools to detect short circuit behavior, confirming its presence in fine-tuned NLU reasoning models. To tackle the short circuit issue, we introduce biologically inspired “crossover” and “mutation” operations. These operations are applied to augment the training data for popular models such as BERT, XLNet, and RoBERTa. Our results demonstrate that these data augmentation techniques effectively enhance the models’ robustness and mitigate the short circuit problem.

Camouflaged object detection is a challenging task that aims to identify objects that are highly similar to their background. Due to the powerful noise-to-image denoising capability of denoising diffusion models, in this paper, we propose a diffusion-based framework for camouflaged object detection, termed diffCOD, a new framework that considers the camouflaged object segmentation task as a denoising diffusion process from noisy masks to object masks. Specifically, the object mask diffuses from the ground-truth masks to a random distribution, and the designed model learns to reverse this noising process. To strengthen the denoising learning, the input image prior is encoded and integrated into the denoising diffusion model to guide the diffusion process. Furthermore, we design an injection attention module (IAM) to interact conditional semantic features extracted from the image with the diffusion noise embedding via the cross-attention mechanism to enhance denoising learning. Extensive experiments on four widely used COD benchmark datasets demonstrate that the proposed method achieves favorable performance compared to the existing 11 state-of-the-art methods, especially in the detailed texture segmentation of camouflaged objects. Our code will be made publicly available later. 

Large-scale pre-trained language models such as BERT have contributed significantly to the development of NLP. However, those models require large computational resources, making it difficult to be applied to mobile devices where computing power is limited. In this paper we aim to address the weakness of existing input-adaptive inference methods which fail to take full advantage of the structure of BERT. We propose Dynamic Routing in BERT, a novel fine-tuning strategy that can accelerate the inference process of BERT through selecting a subsequence of transformer layers list of backbone as a computational path for an input sample. To do this, our approach adds a routing module to the original BERT model to determine whether a layer is included or bypassed during inference. Experimental results on the GLUE benchmark exhibit that our method reduces latency to 75% while maintaining 98% accuracy, yielding a better accuracy-speed trade-off compared to state-of-the-art input-adaptive methods.

Hierarchical Text Classification (HTC) has recently gained traction given the ability to handle complex label hierarchy. This has found applications in domains like E- commerce, Customer care and medicine industry among other real world applications. Existing HTC models either encode label hierarchy separately and mix it with text encoding or guide the label hierarchy structure in the text encoder. Both approaches capture different characteristics of label hierarchy and are complementary to each other. In this paper, we propose a Hierarchical Text Classification using Contrastive Learning Informed Path guided hierarchy (HTC-CLIP), which learns hierarchy-aware text representation and text informed path guided hierarchy representation using contrastive learning. During the training of HTC-CLIP, we learn two different sets of class probabilities distributions and during inference, we use the pooled output of both probabilities for each class to get the best of both representations. Our results show that the two previous approaches can be effectively combined into one architecture to achieve improved performance. Tests on two public benchmark datasets showed an  improvement of 0.99 - 2.37 % in Macro F1 score using HTC-CLIP over the existing state-of-the-art models.

Unsupervised pre-training has demonstrated its potential for accurately constructing world model in visual model-based reinforcement learning (MBRL).  However, most MBRL approaches that rely on unsupervised pre-training exhibit limited generalizability, thereby limiting their practicality in diverse scenarios. These methods produce models that are restricted to the specific task they were trained on, and are not easily adaptable to other tasks. In this work, we introduce a powerful unsupervised pre-training RL framework called VMBRL3, which improves the generalization ability of visual MBRL. VMBRL3 employs task-agnostic videos to pretrain both an autoencoder and world model without access to action or reward information. The fine-tuned world model can then be applied to a range of downstream reinforcement learning tasks, allowing for rapid adaptation to diverse environments and facilitating strategic learning. We demonstrate that our framework significantly improves generalization ability in a variety of manipulation and locomotion tasks. Furthermore, VMBRL3 doubles the sample efficiency and overall performance compared to previous visual RL methods.

Multi-Agent Reinforcement Learning (MARL) has been widely applied in many fields such as smart traffic and unmanned aerial vehicles. However, most MARL algorithms are vulnerable to adversarial perturbations on agent states. Robustness testing for a trained model is an essential step for confirming the trustworthiness of the model against unexpected perturbations. This work proposes a novel Robustness Testing framework for MARL that attacks states of Critical Agents (RTCA). The RTCA has two innovations: 1) a Differential Evolution (DE) based method to select critical agents as victims and to advise the worst-case joint actions on them; and 2) a team cooperation policy evaluation method employed as the objective function for the optimization of DE. Then, adversarial state perturbations of the critical agents are generated based on the worst-case joint actions. This is the first robustness testing framework with varying victim agents. RTCA demonstrates outstanding performance in terms of the number of victim agents and destroying cooperation policies.

We present an approach to non-deterministic planning under full observability via Answer Set Programming. The technique can synthesise compact policies, handle both fair and unfair actions simultaneously, and readily accommodate control knowledge and procedural domain constraints. We show that whereas compact controllers may yield sub-optimal behaviour under a naive executor, optimality can be recovered under a smarter, and still efficient, executor. The planner developed is succinct, elegant, and directly implementable, thus providing higher confidence of its correctness and ease of elaboration. Experimental results show that its performance is competitive against the state-of-the-art.

Contrastive learning has been used to learn useful low-dimensional state representations in visual reinforcement learning (RL). Such state representations substantially improve the sample efficiency of visual RL. Nevertheless, existing contrastive learning-based RL methods have the problem of unstable training. Such instability comes from the fact that contrastive learning requires an extremely large batch size (e.g., 4096 or larger), while current contrastive learning-based RL methods typically set a small batch size (e.g., 512). In this paper, we propose an approach of discrete information bottleneck (DIB) to address this problem. DIB applies the technique of discretization and information bottleneck to contrastive learning in representing the state with concise discrete representation. Using this discrete representation for policy learning results in more stable algorithm training and higher sample efficiency with a small batch size. We demonstrate the advantage of discrete state representation of DIB on several continuous control tasks in the DeepMind Control suite. In the experiments, DIB outperforms prior visual RL methods, both model-based and model-free, in terms of performance and sample efficiency.

In recent years, representations from brain activity patterns and pre-trained language models have been linked to each other based on neural fits to validate hypotheses about language processing. Nonetheless, open questions remain about what intrinsic properties of language processing these neural fits reflect and whether they differ across neural fit approaches, brain networks, and models. In this study, we use parallel sentence and functional magnetic resonance imaging data to perform a comprehensive analysis of four paradigms (masked language modeling, pragmatic coherence, semantic comparison, and contrastive learning) representing linguistic hypotheses about sentence processing. We include three sentence embedding models for each paradigm, resulting in a total of 12 models, and examine differences in their neural fit to four different brain networks using regression-based neural encoding and Representational Similarity Analysis (RSA). Among the different models tested, GPT-2, SkipThoughts, and S-RoBERTa yielded the strongest correlations with language network patterns, whereas contrastive learning-based models resulted in overall low neural fits. Our findings demonstrate that neural fits vary across brain networks and models representing the same linguistic hypothesis (e.g., GPT-2 and GPT-3). More importantly, we show the need for both neural encoding and RSA as complementary methods to provide full understanding of neural fits.

Abusive language detection models tend to have a gender bias problem in which the model is biased towards sentences containing identity words of specific gender groups. Previous studies to reduce bias, such as projection methods, lose information in word vectors and sentence context, reducing detection accuracy. This paper proposes a bias mitigation method that optimizes gender bias mitigation and original information preservation by regularizing sentence embedding vectors based on information theory. Latent vectors generated by an autoencoder are debiased through dual regularization using a gender discriminator, an abuse classifier, and a decoder. While the gender discriminator labels are rumpled, the discriminator softens the gender feature, and the classifier retains the abuse information. Latent vectors are regularized through information theoretic adversarial optimization that disentangles and mitigates gender features. We show that the proposed method successfully orthogonalizes the direction of the correlated information and reduces the gender feature by calculation of subspaces and embedding vector visualization. Moreover, the proposed method maintains the highest accuracy among the four state-of-the-art bias mitigation methods and shows superior performance in reducing gender bias in four different Twitter datasets for abusive language detection.

Self-supervised graph representation learning (SSGRL) is a representation learning paradigm used to reduce or avoid manual labeling. An essential part of SSGRL is graph data augmentation. Existing methods usually rely on heuristics commonly identified through trial and error and are effective only within some application domains. Also, it is not clear why one heuristic is better than another. Moreover, recent studies have argued against some techniques (e.g., dropout: that can change the properties of molecular graphs or destroy relevant signals for graph-based document classification tasks). In this study, we propose a novel data-driven SSGRL approach that automatically learns a suitable graph augmentation from the signal encoded in the graph (i.e., the nodes’ predictive feature and topological information). We propose two complementary approaches that produce learnable feature and topological augmentations. The former learns multi-view augmentation of node features, and the latter learns a high-order view of the topology. Moreover, the augmentations are jointly learned with the representation. Our approach is general that it can be applied to homogeneous and heterogeneous graphs. We perform extensive experiments on node classification (using nine homogeneous and heterogeneous datasets) and graph property prediction (using another eight datasets). The results show that the proposed method matches or outperforms the SOTA SSGRL baselines and performs similarly to semi-supervised methods. The anonymised source code is available at https://anonymous.4open.science/r/dsgrl-D086

Research has shown that cooperative action struggles to emerge in the noisy variant of the donation game, a simple model of noisy multi-agent systems where indirect reciprocity is required to maximise utility. Such noise can arise when agents may have an incorrect view of the reputation of their interaction partners, or when the actions themselves may fail. Concepts such as generosity, as well as the use of higher-order norms, have been investigated as mechanisms to facilitate cooperation in such environments, but often are not effective or require additional assumptions or infrastructure in the system to operate. In this paper, we demonstrate both analytically and empirically that a simple form of generosity when combined with  fine grained reputation can help cooperation emerge. We also show that the use of individual forgiveness strategies rather than the presence of global generosity can support cooperation in such environments.

Implicit imitation assumes that learning agents observe only the state transitions of an agent they use as a mentor, and try to recreate them based on their own abilities and knowledge of their environment. In this paper, we put forward a deep implicit imitation Q-network (DIIQN) model, which incorporates ideas from three well-known Deep Q-Network (DQN) variants. As such, we enable a novel implicit imitation method for online, model-free deep reinforcement learning. Our thorough experimentation in the complex environment of the emerging lane-free traffic paradigm, verifies the benefits of our approach. Specifically, we show that deep implicit imitation RL dramatically accelerates the learning process when compared to a "vanilla" DQN method; and, unlike explicit imitation reinforcement learning, it is able to outperform mentor performance without resorting to additional information, such as the mentor's actions.

We introduce Incomplete Bipolar Argumentation Frameworks (iBAFs), the extension of Dung's Abstract Argumentation Frameworks (AAFs) allowing the simultaneous presence of supports (borrowed from BAFs - Bipolar AAFs) and of uncertain elements of the argumentation graph (borrowed from iAAFs - incomplete AAFs). We investigate the computational complexity of the acceptance problem and the verification problem (under the possible perspective), by performing an analysis of sensitivity to the semantics of supports and the semantics of extensions. On the one hand, we show that adding supports on top of incompleteness does not affect the complexity of the acceptance. On the other hand, surprisingly, we show that the joint use of bipolarity and incompleteness has a deep impact on the verification's complexity: for the Dungean semantics under which the verification over AAFs is polynomial-time solvable, although moving from AAFs to BAFs or to iAAFs does not change the complexity, the verification's complexity over iBAFs may increase up to NP-complete.

Adversarial machine learning (AML), by designing attacks that intentionally break or misuse state-of-the-art machine learning models, has become the most prominent scientific field to explore the security aspects of Artificial Intelligence. A whole range of vulnerabilities, previously irrelevant in traditional ICT, have effectively emerged in these studies. In the light of upcoming legislations mandating security requirements for AI products and services, there is a need to understand how AML techniques connect with the broader field of cybersecurity, and how to articulate more tightly mathematical threat models with realistic cybersecurity procedures. This article aims to contribute to closing the gap between AML and cybersecurity by proposing an approach to study the feasibility of an attack in a cybersecurity risk assessment framework, illustrated with a specific use case of an evasion attack designed to fool traffic sign recognition systems in the physical world. The importance of considering the feasibility of carrying out such attacks under real conditions is emphasized through the analysis of two factors: the reproducibility of the attack according to a published description or existing code, and the applicability of the attack by a malicious actor operating in a real-world environment.

An opinion diffusion scenario is considered where two marketers compete to diffuse their own opinions over a social network. In particular, they implement social proof marketing approaches that naturally give rise to a strategic setting, where it is crucial to find the appropriate order for targeting the individuals to which provide the incentives to adopt their opinions. The setting is extensively studied from the theoretical and empirical viewpoint, by considering strategies defined in a compact way, such as those that can be defined by selecting the individuals according to their degree of centrality in the underlying network. In addition to depicting a clear picture of the complexity issues arising in the setting, several compact strategies are empirically compared on real-world social networks. Results suggest that the effectiveness of compact strategies is moderately influenced by the characteristic of the network, with some centrality measures naturally emerging as good candidates to define heuristic approaches for marketing campaigns. 

Lexical simplification (LS) methods based on pretrained language models have made remarkable progress, generating potential substitutes for a complex word through analysis of its contextual surroundings. However, these methods require separate pretrained models for different languages and disregard the preservation of sentence meaning. In this paper, we propose a novel multilingual LS method via paraphrase generation, as paraphrases provide diversity in word selection while preserving the sentence's meaning. We regard paraphrasing as a zero-shot translation task within multilingual neural machine translation that supports hundreds of languages. After feeding the input sentence into the encoder of paraphrase modeling, we generate the substitutes based on a novel decoding strategy that concentrates solely on the lexical variations of the complex word. Experimental results demonstrate that our approach surpasses BERT-based methods and zero-shot GPT3-based method significantly on English, Spanish, and Portuguese.

Transformer-based pretrained language models (PLMs) have achieved great success in modern NLP. An important advantage of PLMs is good out-of-distribution (OOD) robustness. Recently, diffusion models have attracted a lot of work to apply diffusion to PLMs. It remains under-explored how diffusion influences PLMs on OOD data. The core of diffusion models is a forward diffusion process which gradually applies Gaussian noise to inputs, and a reverse denoising process which removes noise. The noised input reconstruction is a fundamental ability of diffusion models. We directly analyze OOD robustness by measuring the reconstruction loss, including testing the abilities to reconstruct OOD data, and to detect OOD samples. Experiments are conducted by analyzing different training parameters and data statistical features on eight datasets. It shows that finetuning PLMs with diffusion degrades the reconstruction ability on OOD data. The comparison also shows that diffusion models can effectively detect OOD samples, achieving state-of-the-art performance in most of the datasets with an absolute accuracy improvement up to 18%.  These results indicate that diffusion reduces OOD robustness of PLMs.

Deep learning algorithms perform poorly on long-tailed datasets because there is insufficient data in the tail classes to recover its original distribution, resulting in an under-representation of the tail classes in the model. In this work, we propose H2T-FAST, a Head-to-Tail Feature Augmentation method by Style Transfer to improve the performance of the tail. H2T-FAST has the following advantages: (1) It is a fast and universal method that acts on the feature space and so, it can be applied to different backbone networks as well as easily integrated into various imbalanced algorithms with stable performance gains; and (2) it is used only in the training phase and therefore, imposes no additional burden on the deep neural network in the testing phase. In particular, we firstly and randomly select the same number of head samples as the tail ones in each training mini batch. Secondly, the style of the head is transferred to the tail to generate new tail data containing the head style, as a way to increase the number of the tail and get better feature representations. We test our methods on several benchmark vision tasks with state-of-the-art performances.

Link prediction for Knowledge Graphs (KGs) aims to predict missing links between entities. Previous works have utilized Graph Neural Networks (GNNs) to learn specific embeddings of entities and relations. However, these works only consider the linear aggregation of neighbors and do not consider interactions among neighbors, resulting in the neglect of partial indicating information. To address this issue, we propose Deep Interactions-boosted Embeddings (DInBE) which encodes interaction information to enrich the entity representations. To obtain interaction information, we disentangle the representation behind entities to learn diverse disentangled representations for each entity. Then, we learn intra-interactions among neighboring entities in the same component and inter-interactions among different components based on these disentangled representations. With the help of interaction information, our model generates more expressive representations. In addition, we propose a relation-aware scoring mechanism to select useful components based on the given query. Our experiments demonstrate that our proposed model outperforms existing state-of-the-art methods by a large margin in the link prediction task, and this verifies the effectiveness of exploring interactions and adaptive scoring.

Trends and opinion mining in social media increasingly focus on novel interactions involving visual media, like images and short videos, in addition to text. In this work, we tackle the problem of visual sentiment analysis of social media images -- specifically, the prediction of image sentiment polarity. While previous work relied on manually labeled training sets, we propose an automated approach for building sentiment polarity classifiers based on a cross-modal distillation paradigm; starting from scraped multimodal (text + images) data, we train a student model on the visual modality based on the outputs of a textual teacher model that analyses the sentiment of the corresponding textual modality. We applied our method to randomly collected images crawled from Twitter over three months and produced, after automatic cleaning, a weakly-labeled dataset of $\sim$1.5 million images. Despite exploiting noisy labeled samples, our training pipeline produces classifiers showing strong generalization capabilities and outperforming the current state of the art on five manually labeled benchmarks for image sentiment polarity prediction.

Question generation is an important task that helps to improve question answering performance and augment search interfaces with possible suggested questions. While multiple approaches have been proposed for this task, none addresses the goal of generating a diverse set of questions given the same input context. The main reason for this is the lack of multi-reference datasets for training such models. We propose to bridge this gap by seeding a baseline question generation model with named entities as candidate answers. This allows us to automatically synthesize an unlimited number of question-answer pairs. We then propose an approach designed to leverage such multi-reference annotations, and demonstrate its advantages over the standard training and decoding strategies used in question generation. An experimental evaluation on synthetic, as well as manually annotated data shows that our approach can be used in creating a single generative model that produces a diverse set of question-answer pairs per input sentence. 

In recent years, traffic conditions in centralised cities have become more severe. To optimise public resources and reduce congestion, transportation departments rely on traffic prediction. However, unexpected events, for instance, rainfall can impact traffic conditions, which necessitates the introduction of weather elements to improve prediction results. Moreover, most of the existing works characterise the relationship between weather and traffic by simply combining these two issues together. Without carefully designing a structure that captures the inter-dependency between weather and traffic data, it is impossible to produce accurate predictions within a reasonable computational time. To address this issue, we propose a Weather-Aware Graph Attention Network (WeaGAN) that adapts an encoder-decoder architecture with weather attention mechanisms and a gate to model the complex spatial-temporal inter-dependency between weather and traffic adaptively. We further design a self-attention mechanism to improve prediction accuracy. Our experiments on a standard real-world dataset show that, compared to the state-of-the-art, WeaGAN: (i) can improve the prediction accuracy by up to 29\%; and (ii) is efficient in terms of saving up to 61\% of the computation time.

Supervised classification algorithms are used to solve a growing number of real-life problems around the globe. Their performance is strictly connected with the quality of labels used in training. Unfortunately, acquiring good-quality annotations for many tasks is infeasible or too expensive to be done in practice. To tackle this challenge, active learning algorithms are commonly employed to select only the most relevant data for labeling. However, this is possible only when the quality and quantity of labels acquired from experts are sufficient. Unfortunately, in many applications, a trade-off between annotating individual samples by multiple annotators to increase label quality vs. annotating new samples to increase the total number of labeled instances is necessary. In this paper, we address the issue of faulty data annotations in the context of active learning. In particular, we propose two novel annotation unification algorithms that utilize unlabeled parts of the sample space. The proposed methods require little to no intersection between samples annotated by different experts. Our experiments on four public datasets indicate the robustness and superiority of the proposed methods in both, the estimation of the annotator's reliability, and the assignment of actual labels, against the state-of-the-art algorithms and the simple majority voting.

Few-shot relation extraction (FSRE) focuses on detecting new relations through a few annotated instances. Most existing works adopt prototypical network-based models for FSRE. They compute prototype representations for each class in the support set separately, and then use prototype representations for relation prediction. In this way, they learn only the knowledge of each class, regardless of the high-level interactions among these classes. However, these interactions can help the model understand diversity and improve discrimination, which is essential for FSRE, especially for similar relations' prediction. In this work, we introduce a novel Biased Multi-granularity Interaction Prototype Network (BMIPN). Specifically, we mimic human cognitive processes to model explicit and adaptive interactions from intra- and inter-class aspects. Furthermore, we propose a novel biased contrastive learning method that encourages the model to focus on contrasting similar relations, generating discriminative and robust prototype representations. Experimental results on two benchmark datasets demonstrate that BMIPN outperforms state-of-the-art models and achieves better performance with respect to similar relations. 

Federated learning breaks through the barrier of data owners by allowing them to collaboratively train a federated machine learning model without compromising the privacy of their own data. However, Federation Learning also faces the threat of poisoning attacks, especially from the client model updates, which may impair the accuracy of the global model. To defend against the poisoning attacks, previous work aims to identify the malicious updates in high dimensional spaces. However, we find that the distances in high dimensional spaces cannot identify the changes in a small subset of dimensions, and the small changes may affect the global models severely. Based on this finding, we propose an untargeted poisoning attack under the federated learning setting via the partial perturbations on a small subset of the carefully selected model parameters, and present two attack object selection strategies. We experimentally demonstrate that the proposed attack scheme achieves high attack success rate on five state-of-the-art defense schemes. Furthermore, the proposed attack scheme remains effective at low malicious client ratios and still circumvents three defense schemes with a malicious client ratio as low as 2%.

Transformer-based language models demonstrate exceptional performance in Natural Language Processing (NLP) tasks but remain susceptible to backdoor attacks involving hidden input triggers. Trojan injection via hardware bitflips presents a significant challenge for contemporary language models. However, previous research overlooks practical hardware considerations, such as DRAM and cache memory structures, resulting in unrealistic attacks that demand the manipulation of an excessive number of parameters and bits. In this paper, we present TrojBits, a novel approach requiring minimal bit-flips to effectively insert Trojans into real-world Transformer language model systems. This is achieved through a three-module framework designed to efficiently target Transformer-based language models, consisting of Vulnerable Parameters Ranking (VPR), Hardware-aware Attack Optimization (HAO), and Vulnerable Bits Pruning (VBP). Within the VPR module, we are the first to employ Gradient-guided Fisher information to identify the most susceptible Transformer parameters, specifically in the word embedding layer. The HAO module then redistributes these parameters across multiple triggers, conforming to hardware constraints by incorporating a regularization term in the trojan optimization methodology. Finally, the VBP module aims to reduce the number of bit-flips by discarding less significant bits. We evaluate TrojBits on two representative NLP models, BERT and XLNE, on three classification tasks (SST2, OffensEval, and AG's News). Our results demonstrate that our TrojBits successfully achieves the inference-time attack with only 64 parameters out of 116 million and 90-bit flips while maintaining the model performance.

Lifted planning – finding plans directly on the PDDL input model – has attracted renewed attention during the last years. This avoids the process of grounding, which can become computationally prohibitive very easily. However, the main focus of recent research in this area has been on satisficing, i.e., (potentially) suboptimal planning. We present a novel heuristic for optimal lifted planning. Our basic idea is inspired by the LM-cut heuristic, which has been very successful in grounded optimal planning. Like LM-cut, we generate cut-based landmarks via back-chaining from the goal, generating cuts of partially grounded actions. However, exactly mimicking the ground formulation is not feasible, this includes computing the h max heuristic several times for one computation of the LM-cut heuristic (which is already NP-hard to compute). We show that our heuristic is admissible and evaluate it in a cost optimal setting.

For human-like dialogue systems, it is significant to inject the empathetic ability or elicit the opposite's positive emotions, while existing studies mostly only focus on either of the above two research lines. In this work, we propose a novel and grafted task named Empathetic Emotion Elicitation Dialog to make a dialog system able to possess both aspects of ability simultaneously. We do not train an empathetic dialog system and an emotion elicitation dialog system separately and then simply concatenate the responses generated by these two systems, which will cause illogical and repetitive responses. Instead, we propose a unified solution: (1) To generate empathetic responses and emotion elicitation responses within the same semantic space, we design a unified framework. (2) The unified framework has three stages which first retrieve the empathetic and emotion elicitation exemplars as external knowledge, then fine-tune the emotion/action prediction on a pre-trained language model to enhance the empathetic ability, and finally model the user feedback by reinforcement learning to enhance the emotion elicitation ability. Experiments show that our method outperforms the baselines in the response generation quality and simultaneously empathizes with the user and elicits their positive emotions.

This paper is concerned with a refutation system (proof system) for linear constraints, heretofore unaddressed in the literature. Associated with a refutation system are three important features, viz., (a) Soundness,  (b) Completeness, and (c) Efficiency. In this paper, we investigate refutability in Difference Constraint Systems (DCSs) under the Unit Refutation (UR) system. Recall that a difference constraint is a linear relationship of the form: $x_{i}-x_{j} \le b_{ij}$ and a DCS is a conjunction of such constraints. The UR refutation system is {\bf incomplete}, in that unsatisfiable DCSs may not have unit refutations. We establish that this refutation system is efficient in that there exists a tractable algorithm for determining if a DCS has an UR. Investigating {\bf  weak} (incomplete) refutation systems leads to a better understanding of the inference rules required for establishing contradictions in the given constraint system. Thus, this study is well-motivated. Difference constraints are used in the field of abstract interpretation.  Abstract interpretation is an extremely useful technique for program verification. Note that a program can be modeled using a system of linear constraints. Depending on how the model is constructed, a refutation of such a system corresponds to proof that the original program is incorrect. Despite the fact that unit refutations can be exponentially long in terms of the input system size, we provide a compact representation of these refutations. This compact representation is a unique contribution of this paper.

We study Linear Temporal Logic Modulo Theories over Finite Traces (LTLfMT), a recently introduced extension of LTL over finite traces (LTLf) where propositions are replaced by first-order formulas. In general, LTLfMT was shown to be semi-decidable for any decidable first-order theory (e.g., linear arithmetics), with a tableau-based semi-decision procedure. In this paper we focus on theories that satisfy the property of model completion, equivalent to the existence of uniform interpolants. This includes a variety of theories, including arithmetics, theories underlying database systems, and combinations thereof. We show that those theories admit a sound and complete pruning rule for the LTLfMT tableau that, when coupled with additional requirements, guarantees termination as well. This technique allows us to establish new decidability results for several fragments of LTLfMT, as well as to give new decidability proofs for classes that are already known.

Deep neural networks are widely used in image classification tasks, but their internal decision-making mechanisms are often difficult to explain. While various algorithms have been developed to visualize these mechanisms, many of them produce coarse, noisy results that are not always convincing. To address this issue, we propose a method for enhancing saliency maps produced by saliency methods. Our method uses a fixed-size sliding window to upsample local regions of the input image and feed them into the selected visualization algorithm to generate class-specific saliency maps and probability scores. We then downsample the resulting saliency maps and multiply them by the probability scores to obtain maps with greater detail. We evaluate our method using different saliency methods and network architectures, and demonstrate its effectiveness through both quantitative metrics and intuitive evaluation. Our results show that our method significantly improves the performance of these saliency methods, providing a more valid and reliable means of visualizing the decision mechanisms of deep neural networks.

Neural networks (NNs) have various applications in AI, but explaining their decisions remains challenging. Existing approaches often focus on explaining how changing individual inputs affects NNs' outputs. However, an explanation that is consistent with the input-output behaviour of an NN is not necessarily faithful to the actual mechanics thereof. In this paper, we exploit relationships between multi-layer perceptrons (MLPs) and quantitative argumentation frameworks (QAFs) to create argumentative explanations for the mechanics of MLPs. Our SpArX method first sparsifies the MLP while maintaining as much of the original structure as possible. It then translates the sparse MLP into an equivalent QAF to shed light on the underlying decision process of the MLP, producing global and/or local explanations. We demonstrate experimentally that SpArX can give more faithful explanations than existing approaches, while simultaneously providing deeper insights into the actual reasoning process of MLPs.

OWL and SHACL are two prominent W3C standards for managing RDF graphs, the data model of the Web. They are used for different purposes and make different assumptions about the completeness of data: SHACL is used for expressing integrity constraints on complete data, while OWL allows inferring implicit facts from incomplete data; SHACL reasoners perform validation, while OWL reasoners do logical inference. Integrating these two tasks into one
uniform approach is a relevant but challenging problem. The SHACL standard envisions graph validation in combination with OWL entailment, but it does not provide technical guidance on how to realize this. To address this problem, we propose new intuitive semantics for validating SHACL constraints with OWL 2 QL ontologies based on a suitable notion of the chase. We propose an algorithm that rewrites a set of recursive SHACL constraints (with stratified negation) and
an OWL 2 QL ontology into a stand-alone set of SHACL constraints that preserves validation for every input graph, which can in turn be evaluated using an off-the-shelf SHACL validator. We show that validation in this setting is ExpTime complete in combined complexity, but only PTime complete in data complexity, i.e., if the constraints and the ontology are fixed.

Many decision-making scenarios, e.g., public policy, healthcare, business, and disaster response, require accommodating the preferences of multiple stakeholders. We offer, to the best of our knowledge, the first formal treatment of reasoning with multi-stakeholder qualitative preferences in a setting where each stakeholder expresses their preferences in a qualitative preference language, e.g., CP-net, CI-net, TCP-net, CP-Theory. We introduce a query language for expressing queries against preferences of multiple stakeholders over sets of outcomes that satisfy  specified conditions, e.g., $\mlangpref{\psi_1}{\psi_2}{A}$ (read as the set of stakeholders in $A$ prefer outcomes that satisfy $\psi_1$ over outcomes that satisfy $\psi_2)$, etc.  Motivated by practical application scenarios, we introduce and compare alternative semantics for such queries, and examine their interrelationships.  We provide a provably correct algorithm for answering multi-stakeholder preference queries using model checking in alternation-free $\mu$-calculus. The results of our experiments demonstrate the feasibility of our approach.

The dynamics of real-world applications and systems require efficient methods for improving infeasible solutions or restoring corrupted ones by making modifications to the current state of a system in a restricted way. We propose a new framework of solution discovery via reconfiguration for constructing a feasible solution for a given problem by executing a sequence of small modifications starting from a given state. Our framework integrates different aspects of classical local search, reoptimization, and combinatorial reconfiguration. We exemplify our framework on a multitude of fundamental combinatorial problems, namely Vertex Cover, Independent Set, Dominating Set, and Coloring. We study the classical as well as the parameterized complexity of the solution discovery variants of those problems and explore the boundary between tractable and intractable instances.

Dung’s Argumentation Framework (AF) has been extended in several directions. An interesting extension, among others, is the Epistemic AF (EAF) which allows representing the agent’s belief by means of epistemic constraints. In particular, an epistemic constraint is a propositional formula over labeled arguments (e.g. in(a), out(c)) extended with the modal operators K and M that intuitively state that the agent believes that a given formula is certainly or possibly true, respectively. In this paper, focusing on EAF, we investigate the complexity of the possible and necessary variants of three canonical problems in abstract argumentation: verification, existence, and non-empty existence. Moreover, we explore the relationship between EAF and incomplete AF (iAF), an extension of AF where arguments and attacks may be uncertain. Our complexity analysis shows that the verification problem in iAF can be naturally reduced to the verification in EAF, while it turns out that a similar result cannot hold for the necessary (non-empty) existence problem.

We consider ""bidding games"", a class of two-player zero-sum ""graph games"". The game proceeds as follows. Both players have bounded budgets. A token is placed on a vertex of a graph, in each turn the players simultaneously submit bids, and the higher bidder moves the token. In case of ties, Player 1 moves the token. Player 1 wins the game iff the token visits a designated target vertex. We consider, for the first time, ""poorman discrete-bidding"" in which the granularity of the bids is restricted and the higher bid is paid to the bank. In contrast, in  ""Richman continuous-bidding"" bids can be arbitrarily small and are paid to the other player. While the latter mechanism is technically more accessible, the former is more appealing from a practical standpoint. Our study focuses on threshold budgets, which is the necessary and sufficient initial budget required for Player 1 to ensure winning against a given Player 2 budget. We first show existence of thresholds. In DAGs, we show that threshold budgets can be approximated with error bounds by thresholds under continuous-bidding and that they exhibit a periodic behavior. We identify closed-form solutions in special cases. We implement and experiment with an algorithm to find threshold budgets.

We define contrastive explanations that are suited to tree-based classifiers. In our framework, contrastive explanations are based on the set of (possibly non-independent) Boolean characteristics used by the classifier and are at least as general as contrastive explanations based on the set of characteristics of the instances considered at start. We investigate the computational complexity of computing contrastive explanations for any Boolean classifier (including tree-based ones), when the Boolean conditions used are not independent. Finally, we present and evaluate empirically an algorithm for computing minimum-size contrastive explanations for random forests.

The use of Dynamic Epistemic Logic (DEL) in multi-agent planning has led to a widely adopted action formalism that can handle nondeterminism, partial observability and arbitrary knowledge nesting. As such expressive power comes at the cost of undecidability, several decidable fragments have been isolated, mainly based on syntactic restrictions of the action formalism. In this paper, we pursue a novel semantic approach to achieve decidability. Namely, rather than imposing syntactical constraints, the semantic approach focuses on the axioms of the logic for epistemic planning. Specifically, we augment the logic of knowledge S5n and with an interaction axiom called (knowledge) commutativity, which controls the ability of agents to unboundedly reason on the knowledge of other agents. We then provide a threefold contribution. First, we show that the resulting epistemic planning problem is decidable. In doing so, we prove that our framework admits a finitary non-fixpoint characterization of common knowledge, which is of independent interest. Second, we study different generalizations of the commutativity axiom, with the goal of obtaining decidability for more expressive fragments of DEL. Finally, we show that two well-known epistemic planning systems based on action templates, when interpreted under the setting of knowledge, conform to the commutativity axiom, hence proving their decidability.

We investigate winner determination for two popular proportional representation systems: the Monroe and Chamberlin-Courant (abbrv. CC) systems. Our study focuses on (nearly) single-peaked resp. single-crossing preferences. We show that for single-crossing approval preferences, winner determination of the Monroe rule is polynomial, and for both rules, winner determination mostly admits FPT algorithms with respect to the number of voters to delete to obtain single-peaked or single-crossing preferences. Our results answer some complexity questions from the literature [18, 28, 21].

Although current summarization models can process increasingly long text sequences, they still struggle to capture salient related information spread across the extremely-lengthy size of inputs with few labeled training instances. Today's research still relies on standard input truncation without considering graph-based modeling of multiple semantic units to summarize only crucial facets. This paper proposes G-Seek, a graph-based summarization of extracted essential knowledge. By representing the long source with a heterogeneous graph, our method extracts and provides salient sentences to an abstractive summarization model to generate the summary. Experimental results in low-resource scenarios, distinguished by data scarcity, reveal that G-Seek consistently improves both long and multi-document summarization performance and accuracy across several datasets. 

The Vision Transformer architecture has shown to be competitive in the computer vision (CV) space where it has dethroned convolution-based networks in several benchmarks. Nevertheless, convolutional neural networks (CNN) remain the preferential architecture for the representation module in reinforcement learning. In this work, we study pretraining a Vision Transformer using several state-of-the-art self-supervised methods and assess the quality of the learned representations. To show the importance of the temporal dimension in this context we propose an extension of VICReg to better capture temporal relations between observations by adding a temporal order verification task. Our results show that all methods are effective in learning useful representations and avoiding representational collapse for observations from the Atari Learning Environment (ALE) which leads to improvements in data efficiency when we evaluated in reinforcement learning (RL). Moreover, the encoder pretrained with the temporal order verification task shows the best results across all experiments, with richer representations, more focused attention maps and sparser representation vectors throughout the layers of the encoder, which shows the importance of exploring such similarity dimension. With this work, we hope to provide some insights into the representations learned by ViT during a self-supervised pretraining with observations from RL environments and to understand which properties arise in the representations that lead to the best-performing agents.

Our Collaborative Research Center (CRC) aims to enable robots to perform environmental interaction tasks with close to human capacity. It therefore employs a shared ontology to model the activity of both kinds of agents, empowering robots to learn from human experiences. To properly describe these human experiences, the ontology will strongly benefit from incorporating characteristics of neuronal information processing which are not accessible from a behavioral perspective alone. We, therefore, propose the analysis of human neuroimaging data for evaluation and validation of concepts and events defined in the ontology model underlying most of the CRC projects. In an exploratory analysis, we employed an Independent Component Analysis (ICA) on functional Magnetic Resonance Imaging (fMRI) data from participants who were presented with the same complex video stimuli of activities as robotic and human agents in different environments and contexts. We then correlated the activity patterns of brain networks represented by derived components with timings of annotated event categories as defined by the ontology model. The present results demonstrate a subset of common networks with stable correlations and specificity towards particular event classes and groups, associated with environmental and contextual factors. These neuronal characteristics will open up avenues for adapting the ontology model to be more consistent with human information processing.

We introduce Strategy Repair, the problem of finding a minimal amount of modifications to turn a strategy for a reachability game from losing into winning. The problem is relevant for a number of settings in Planning and Synthesis, where solutions essentially correspond to winning strategies in a suitably defined reachability game. We show, via reduction from Vertex Cover, that Strategy Repair is NP-complete and devise two algorithms, one exact and exponential and one polynomial but suboptimal, which we compared experimentally. The reported experimentation includes some heuristics for strategy modification, which proved crucial in dramatically improving performance.

Recent work demonstrated that path explanation redundancy is ubiquitous in decision trees, i.e. most often paths in decision trees include literals that are redundant for explaining a prediction. The implication of this result is that decision trees must be explained. Nevertheless, there are applications of DTs where running an explanation algorithm is impractical. For example, in settings that are time or power constrained, running software algorithms for explaining predictions would be undesirable. Although the explanations for paths in DTs do not generally represent themselves a decision tree, this paper shows that one can construct a decision set from some of the decision tree explanations, such that the decision set is not only explained, but it also exhibits a number of properties that are critical for replacing the original decision tree.

In numeric AI planning, a state is represented by propositions and numeric variables, actions change the values of numeric variables in addition to adding and deleting propositions, and goals and preconditions of actions may include conditions over numeric variables.  While domains of numeric variables are rational numbers in general, upper and lower bounds on variables affected only by constant increase and decrease can sometimes be determined and exploited by a heuristic function. In this paper, we generalize the existing method to variables that are changed by linear effects. We exploit the extracted bounds to improve the numeric LM-cut heuristic, a state-of-the-art admissible heuristic for linear numeric planning. Empirical evaluation shows that our method improves the performance of LM-cut in multiple domains. The proposed method can also detect unsolvability of some numeric tasks in polynomial time.

Disjunctive scheduling problems such as the job shop and open shop are at the heart of many real world scheduling instances. In this paper, we frame such problems as disjunctive temporal networks associated with a makespan minimization objective. For those, we propose a hybrid approach between SMT and CP solvers. In particular, we keep from SMT solvers the aggregated constraint propagation in decision procedures as well as the explanations and clause learning mechanisms upon conflict. However, like all CP solvers, we maintain an explicit domain representation of integer variables, tightly integrated with clause learning. Automated search exploits explanations to derive activity-based heuristics combined with more classical value-based heuristics of CP solvers. The resulting solver is compared to state-of-the-art exhaustive search solvers on the classical benchmarks of job shop and open shop problems.

A well-defined reward function is crucial for successful training of an reinforcement learning (RL) agent. However, defining a suitable reward function is a notoriously challenging task, especially in complex, multi-objective environments. To define a reward function in such environments an initial, potentially misspecified reward function is often iteratively adjusted based on observed learned behavior. In this work, we aim to automate this process by proposing ITERS, an iterative reward shaping approach using human feedback for mitigating the effects of a misspecified reward function. Our approach allows the user to provide trajectory-level feedback on agent's behavior during training, which can be integrated as a reward shaping signal in the following training iteration. We also allow the user to provide explanations of their feedback, which are used to augment the feedback and reduce user effort and feedback frequency. We evaluate ITERS in three environments and show that it can successfully correct misspecified reward functions. 

Explainability is crucial for the successful use of AI for fake news detection (FND). Researchers aim to improve the explainability of FND by highlighting important descriptions in crowd-contributed comments as clues. From the perspective of law and sociology, there are distinct clues that are easy to discover and understand, and faint clues that require careful observation and analysis. For example, in fake new related to COVID-Omicron showing increased pathogenicity and transmissibility, distinct clues might involve virologists’ opinions regarding the inverse correlation between pathogenicity and transmissibility. Meanwhile, faint clues might be reflected in an infected person’s claim that the symptoms are milder than a cold (indirectly indicating reduced pathogenicity). Occasionally, the statements of some ordinary eyewitness can decisively reveal the truth of news, leading to the judgment of fake news. Existing methods generally use static networks to model the entire news life-cycle, which makes it fail to capture the subtle dynamic interactions between individual clues and news. Thereby faint clues, whose relations to the truth of news are challenging to be characterized and extracted directly, are more likely to be overshadowed by distinct clues. To address this issue, we propose an explainable FND method, dubbed as PiercingEye, which leverages dynamic interaction information to progressively mine valuable clues. PiercingEye models the news propagation topology as a dynamic graph, with interactive comments serving as nodes, and employs the time-semantic encoding mechanism to refine the modeling of temporal interaction information between comments and news to preserve faint clues. Subsequently, it utilizes the self-attention mechanism to aggregate distinct and faint clues for FND. Experimental results demonstrate that PiercingEye outperforms state-of-the-art methods and is capable of identifying both faint and distinct clues for humans to debunk fake news.

In a multi-agent system, agents share their local observations to gain global situational awareness for decision making and collaboration using a message passing system. When to send a message, how to encode a message, and how to leverage the received messages directly affect the effectiveness of the collaboration among agents. When training a multi-agent cooperative game using reinforcement learning (RL), the message passing system needs to be optimized together with the agent policies. This consequently increases the model’s complexity and poses significant challenges to the convergence and performance of learning. To address this issue, we propose the Belief-map Assisted Multi-agent System (BAMS), which leverages a neuro-symbolic belief map to enhance training. The belief map decodes the agent's hidden state to provide a symbolic representation of the agent's understanding of the environment and other agents' status.  he simplicity of symbolic representation allows the gathering and comparison of the ground truth information for the belief, which provides an additional channel of feedback for the learning. Compared to the sporadic and delayed feedback coming from the reward in RL, the feedback from the belief map is more consistent and reliable. Agents using BAMS can learn a more effective message passing network to better understand each other, resulting in better performance in the game. We evaluate BAMS's performance in a cooperative predator and prey game with varying levels of map complexity and compare it to previous multi-agent message passing models. The simulation results showed that BAMS reduced training epochs by 66%, and agents who apply the BAMS model completed the game with 34.62% fewer steps on average.

Offline reinforcement learning (RL) aims to learn effective policies from recorded data without further interactions in the environments that are often costly or risky. Model-based algorithms, which begin by constructing an environmental model and then learn the policy under the model, have become a promising approach. However, most existing works have been over-conservative to avoid the out-of-distribution error induced by the model generated samples, leading to poor performance instead. In this work, we propose a novel model-based offline RL method, named Simple Double Validation (SDV). The main idea of SDV is to introduce an additional guidance model to assist the agent in determining the rationality of the states, combined with an advantage weighting factor to avoid effects that could potentially mislead the models due to suboptimal samples. In this way, the agent can be guided to more favourable states with reliable decisions. We evaluated SDV on the widely studied offline RL benchmarks and demonstrated its state-of-the-art performance. At the same time, our work introduces the idea of double validation and model advantage weighting into the field of model-based offline RL, providing new insights for future research.

Modern recommender systems utilize users' historical behaviors to generate personalized recommendations. However, these systems often lack user controllability, leading to diminished user satisfaction and trust in the systems. Acknowledging the recent advancements in explainable recommender systems that enhance users' understanding of recommendation mechanisms, we propose leveraging these advancements to improve user controllability. In this paper, we present a user-controllable recommender system that seamlessly integrates explainability and controllability within a unified framework. By providing both retrospective and prospective explanations through counterfactual reasoning, users can customize their control over the system by interacting with these explanations. Furthermore, we introduce and assess two attributes of controllability in recommendation systems: controllability complexity and controllability accuracy. Experimental evaluations on MovieLens and Yelp datasets substantiate the effectiveness of our proposed framework. Additionally, our experiments demonstrate that offering users control options can potentially enhance recommendation accuracy in the future.

Safety is critical to broadening the application of reinforcement learning (RL). Often, we train RL agents in a controlled environment, such as a laboratory, before deploying them in the real world. However, the real-world target task might be unknown prior to deployment. Reward-free RL trains an agent without the reward to adapt quickly once the reward is revealed. We consider the constrained reward-free setting, where an agent (the guide) learns to explore safely without the reward signal. This agent is trained in a controlled environment, which allows unsafe interactions and still provides the safety signal. After the target task is revealed, safety violations are not allowed anymore. Thus, the guide is leveraged to compose a safe behavior policy. Drawing from transfer learning, we also regularize a target policy (the student) towards the guide while the student is unreliable and gradually eliminate the influence from the guide as training progresses. The empirical analysis shows that this method can achieve safe transfer learning and helps the student solve the target task faster.

There have been a number of attempts to develop a formal definition of causality that accords with our intuitions about what constitutes a cause. Perhaps the best known is the "modified'' definition of actual causality, HPm, due to Halpern. In this paper, we argue that HPm gives counterintuitive results for some simple causal models. We propose Dynamic Causality (DC), an alternative semantics for causal models that leads to an alternative definition of causes. DC ascribes the same causes as HPm on the examples of causal models widely discussed in the literature and ascribes intuitive causes for the kinds of causal models we consider. Moreover, we show that the complexity of determining a cause under the DC definition is lower than for the HPm definition.

Recently, there has been a growing interest in developing machine learning (ML) models that can promote fairness, i.e., eliminating biased predictions towards certain populations (e.g., individuals from a specific demographic group). Most existing works learn such models based on well-designed fairness constraints in optimization. Nevertheless, in many practical ML tasks, only very few labeled data samples can be collected, which can lead to inferior fairness performance. This is because existing fairness constraints are designed to restrict the prediction disparity among different sensitive groups, but with few samples, it becomes difficult to accurately measure the disparity, thus rendering ineffective fairness optimization. In this paper, we define the fairness-aware learning task with limited training samples as the fair few-shot learning problem. To deal with this problem, we devise a novel framework that accumulates fairness-aware knowledge across different meta-training tasks and then generalizes the learned knowledge to meta-test tasks via fairness adaptation. To compensate for insufficient training samples, we propose an essential strategy to select and leverage an auxiliary set for each meta-test task. These auxiliary sets contain several labeled training samples that can enhance fairness adaptation in meta-test tasks, thereby allowing for the transfer of learned useful fairness-oriented knowledge to meta-test tasks. Furthermore, we conduct extensive experiments on three real-world datasets to validate the superiority of our framework against the state-of-the-art baselines.

Recent studies show that an ensemble of deep networks can have better adversarial robustness by increasing the learning diversity of base models to limit adversarial transferability. However, existing schemes mostly rely on a second-order method for gradient regularization which usually involves a heavy computation overhead. In this paper, we propose a simple yet effective method which eliminates the use of a second-order optimization and significantly reduces the computation complexity of regularized simultaneous training of deep ensemble networks. For the first time, we show analytically that stochastic regularization by the proposed approach can promote both model smoothness and feature diversity of representation learning in the deep space. We also show that the proposed method is able to achieve a better gain of certified robustness. This is due to the effect of a prioritized feature selection enabled by an adaptive and continuous sampling of neuron activation among the base networks. Experimental results show that our method can improve adversarial robustness significantly comparing with the existing ensemble models on several image benchmark datasets. The ensemble performance can be further boosted by complementing the stochastic regularization approach with other defense paradigms such as adversarial training.

This research proposes an interpretable and practical solution for learning the causal relationship between a desired treatment and its outcome from observed data. This is crucial to assist decision-makers in various fields. However, in real-world scenarios, learning causal relationships from observed data is difficult due to confounding factors that influence treatment and outcome, and are not reported for all treatments, a problem known as Selective Confounding. This issue is prevalent in child welfare, where caseworkers determine whether children should remain in their family homes or be assigned out-of-home placements like foster care during child maltreatment assessments. In such cases, factors reported for children kept with their family are often lower than those taken out of home, making it difficult to observe all confounding factors under desired treatment. Our proposed scheme uses dual-treatment samples to identify the target quantity under Selective Confounding and adapts two-step procedures to learn counterfactual predictors such as Regression Adjustment or Doubly-Robust. We provide both theoretical error bounds and empirical evidence of the effectiveness of our proposed scheme using synthetic and real-world child placement data. Additionally, we propose three evaluation methods specifically tailored to child placement data. By promoting transparency and interpretability, we aim to provide a valuable tool for decision-makers.

Significant advances have been recently made in the development of increasingly effective in-exact (or incomplete) search algorithms---particularly geared towards finding good though not provably optimal solutions fast---for the constraint optimization paradigm of maximum satisfiability (MaxSAT). One of the most successful recent approaches is a new type of stochastic local search in which---to an extent counterintuitively---a Boolean satisfiability (SAT) solver is used as a decision oracle for moving from a solution to another. In this work, we strive for extending the success of the approach to the more general realm of pseudo-Boolean optimization (PBO), where constraints are expressed as linear inequalities over binary variables. As a basis for the approach, we make use of recent advances in practical approaches to satisfiability checking pseudo-Boolean constraints. We outline various heuristics within the oracle-based approach to anytime PBO solving, and show that the approach compares in practice favourably both to a recently-proposed local search approach for PBO that is in comparison a more traditional instantiation of the stochastic local search paradigm, and a recent exact PBO approach when used as an anytime solver.   

Emotion recognition in conversations (ERC) typically involves modeling both intra- and inter-speaker context dependencies. While modeling inter-speaker dependencies has limitations in highlighting differences among listeners. Recent research in ERC has focused on incorporating causal knowledge to enhance utterance representation. However, the incorporation of causal knowledge ignores the consistency of speakers' causal cues (SCC) in modeling speaker-level context, which does not meet emotional dynamics in conversations. Additionally, it is observed that historical utterances from various topics are blindly leveraged in dialogue context modeling, which fails the inter- and intra-topic coherence. To address these issues, we propose the topic- and causal-aware graph interactive network (TCA-GIN). Specifically, we suggest a graph encoder to model topic-level context dependencies based on a context-sensitive neural topic model, achieving inter- and intra-topic coherences. Then, we present a causal-aware graph attention to keep the consistency of SCC, improving speaker-level context modeling. Finally, considering the defect of context modeling insensitive to the differences among inter-speaker or inter-topic dependencies, we employ supervised comparative learning to sweeten it. Experimental results show that TCA-GIN outperforms state-of-the-art methods on three public conversational datasets.

Current AI regulations require discarding sensitive features (e.g., gender, race, religion) in the algorithm's decision-making process to prevent unfair outcomes. However, even without sensitive features in the training set,  algorithms can persist in discrimination. Indeed, when sensitive features are omitted (fairness under unawareness), they could be inferred through non-linear relations with the so-called proxy features. In this work, we propose a way to reveal the potential hidden bias of a machine learning model that can persist even when sensitive features are discarded. This study shows that it is possible to unveil whether the black-box predictor is still biased by exploiting counterfactual reasoning. In detail, when the predictor provides a negative classification outcome, our approach first builds counterfactual examples for a discriminated user category to obtain a positive outcome. Then, the same counterfactual samples feed an external classifier (that targets a sensitive feature) that reveals whether the modifications to the user characteristics needed for a positive outcome moved the individual to the non-discriminated group. When this occurs, it could be a warning sign for discriminatory behavior in the decision process. Furthermore, we leverage the deviation of counterfactuals from the original sample to determine which features are proxies of specific sensitive information. Our experiments show that, even if the model is trained without sensitive features, it often suffers discriminatory biases.

Training reinforcement learning (RL) agents using scalar reward signals is often infeasible when an environment has sparse and non-Markovian rewards. Moreover, handcrafting these reward functions before training is prone to misspecification. We learn non-Markovian finite task specifications as finite-state `task automata' from episodes of agent experience within environments with unknown dynamics. First, we learn a product MDP, a model composed of the specification's automaton and the environment's MDP (both initially unknown), by treating it as a partially observable MDP and employing hidden Markov model learning algorithms. Second, we efficiently distil the task automaton (assumed to be a deterministic finite automaton) from the learnt product MDP. Our automaton enables a task to be decomposed into sub-tasks, so an RL agent can later synthesise an optimal policy more efficiently. It is also an interpretable encoding of high-level task features, so a human can verify that the agent's learnt tasks have no misspecifications. We also take steps towards ensuring that the automaton is environment-agnostic, making it well-suited for use in transfer learning. 

Large-scale language models such as DNABert and LOGO aim to learn optimal gene representations and are trained on the entire Human Reference Genome. However, standard tokenization schemes involve a simple sliding window of tokens like k-mers that do not leverage any gene-based semantics and thus may lead to (trivial) masking of easily predictable sequences, and subsequently inefficient Masked Language Modeling (MLM) training. Therefore, we propose a novel masking algorithm, GeneMask, for MLM training of gene sequences, where we randomly identify positions in a gene sequence as mask centers and locally select the span around the mask center with the highest Normalized Pointwise Mutual Information (NPMI) to mask. We observe that in the absence of human-understandable semantics in the genomics domain (in contrast, semantic units like words and phrases are inherently available in NLP). GeneMask-based models substantially outperform the SOTA models (DNABert and LOGO) over four benchmark gene sequence classification datasets in five few-shot settings (10 to 1000-shot). More significantly, the GeneMask-based DNABert model is trained for less than one-tenth of the number of epochs of the original SOTA model. We also observe a strong correlation between top-ranked PMI tokens and conserved DNA sequence motifs, which may indicate the incorporation of latent genomic information. The codes (including trained models) and datasets are made publicly available at https://github.com/clinical-trial/GeneMask.

The use of multi-queue heuristic search and tie-breaking strategies has shown to be very effective for satisficing planning in the Classical Planning setting. However, to the best of our knowledge, the use of such techniques has never been studied and employed in heuristic search algorithms for Fully Observable Non-Deterministic (FOND) Planning. In this paper, we adapt existing satisficing techniques for scaling-up an AND/OR heuristic search algorithm for FOND Planning. Namely, we employ multi-queue heuristic search, dead-end detection, and tie-breaking strategies in LAO* for improving the extraction of strong-cyclic policies. We assess the efficiency of our alternative techniques in LAO* through an extensive ablation study over two different FOND Planning benchmarks. Empirical results show that our techniques improve the performance of LAO* in terms of coverage, expanded nodes, and planning time compared to a well-known planner based on vanilla LAO*. Indeed, the best configuration of our techniques is competitive with the current state-of-the-art in FOND Planning.

One of the main appeals of AlphaZero-style game-playing agents, which combine deep learning and Monte Carlo Tree Search, is that they can be trained autonomously without external expert-level domain knowledge. However, training such agents is generally computationally expensive, with the most computationally time-consuming step being generating training data via self-play. Here we propose an improved strategy for generating self-play training data, resulting in higher-quality samples, especially in earlier training phases. The new strategy initially emphasizes the latter game phases and gradually extends those phases to entire games as the training progresses. In our test domains, the games Connect4 and Breakthrough, we show that game-playing agents using the improved training approach learn significantly faster than counterpart agents using a standard approach. Furthermore, we empirically show that the proposed strategy is (in our test domains) superior to several recently proposed strategies for expediting self-play learning in game playing.

Many decision problems cannot be solved exactly and use several estimation algorithms that assign scores to the different available options. The estimation errors can have various correlations, from low (e.g. between two very different approaches) to high (e.g. when using a given algorithm with different hyperparameters). Most aggregation rules would suffer from this diversity of correlations. In this article, we propose different aggregation rules that take correlations into account, and we compare them to naive rules in various experiments based on synthetic data. Our results show that when sufficient information is known about the correlations between errors, a maximum likelihood aggregation should be preferred. Otherwise, typically with limited training data, we recommend a method that we call Embedded Voting (EV).

In the roommate problem, pairs of agents must be formed, based on ordinal preferences of the agents over each other. In this article, we examine fair roommate matchings by relaxing envy-freeness to account for justified envy based on the rank in the agents’ preferences. A rank-envy-free matching prevents that an agent prefers the partner of another agent whereas she has ranked it better. Although this requirement is pretty weak in house allocation [Belahcène et al., 2021], we show that it is more demanding in the roommate setting. We study parameterizations of rank-envy-freeness, as well as further natural relaxations of this concept. We also investigate the connections between this family of rank-based fairness criteria and known optimality or stability concepts.

Although machine learning models achieve high classification accuracy against benign samples, they are vulnerable to adversarial machine learning (AML) attacks which generate adversarial samples by adding well-crafted perturbations to the benign samples. The perturbations can be increased to enhance the attack success rate, however, if the perturbations are added without considering the semantic or perceptual similarity between the benign and adversarial samples, the attack can be easily perceived/detected. As such, there exists a trade-off between the attack success rate and the perceptual similarity. In this paper, we propose a novel Semantic-Preserving Adversarial Transformation (SPAT) framework which facilitates an advantageous trade-off between the two metrics. SPAT modifies the optimisation objective of an AML attack to include the goal of increasing the attack success rate as well as the goal of maintaining the perceptual similarity between benign and adversarial samples. Our experiments on a variety of datasets including CIFAR-10, GTSRB, and MNIST demonstrate that SPAT-transformed AML attacks achieve higher attack success rates than the conventional AML attacks while maintaining the same perceptual similarity.

Counterfactual Explanations (cf-XAI) describe the smallest changes in feature values necessary to change an outcome from one class to another. However, presently many cf-XAI methods neglect the feasibility of those changes. In this paper, we introduce a novel approach for presenting cf-XAI in natural language (Natural-XAI), giving careful consideration to actionable and comprehensible aspects while remaining cognizant of immutability and ethical concerns. We present three contributions to this endeavor. Firstly, through a user study, we identify two types of themes present in cf-XAI composed by humans: content-related, focusing on how features and their values are included from both the counterfactual and the query perspectives; and structure-related, focusing on the structure and terminology used for describing necessary value changes. Secondly, we introduce a feature actionability taxonomy with four clearly defined categories, each accompanied by an example, to streamline the explanation presentation process. Using insights from the user study and our taxonomy, we created a generalisable template-based natural language generation (NLG) method compatible with existing explainers like DICE, NICE, and DisCERN, to produce counterfactuals that address the aforementioned limitations of existing approaches. Finally, we conducted a second user study to assess the performance of our taxonomy-guided NLG templates on three domains. Our findings show that the taxonomy-guided Natural-XAI approach (n-XAI^T ) received higher user ratings across all dimensions, with significantly improved results in the majority of the domains assessed for articulation, acceptability, feasibility, and sensitivity dimensions.

Multi-class imbalance learning usually confronts more challenges especially when learning from streaming data. Most existing methods focus on manipulating class imbalance ratios, disregarding other data properties such as the borderline and the disjunct. Recent studies have shown non-negligible impact of disregarding these properties on deteriorating predictive performance. Online multi-class imbalance would further exacerbate such negative impact. To abridge the research gap of online multi-class imbalance learning, we propose to enhance the number of training times of borderline samples based on the disjunct class-wise clusters that are adaptively constructed over time for each class individually. Specifically, we propose a borderline enhanced strategy for ensemble aiming to increase the number of training times of samples neighboring to borderline areas of different classes. We also propose to generate synthetic samples for training based on the adaptively learned disjunct clusters that are maintained for each class individually online, catering for online multi-class imbalance problem directly. These two components construct the Borderline Enhanced Disjunct Cluster Based Oversampling Ensemble (BEDCOE). Experimental studies are conducted to investigate the effectiveness of BEDCOE and each of its components in dealing with online multi-class imbalance.

Matching problems with group-fairness constraints and diversity constraints have numerous applications such as in allocation problems, committee selection, school choice, etc. Moreover, online matching problems have lots of applications in ad allocations and other e-commerce problems like product recommendation in digital marketing. We study two problems involving assigning items to platforms, where items belong to various groups depending on their attributes; the set of items are available offline and the platforms arrive online. In the first problem, we study online matchings with proportional fairness constraints. Here, each platform on arrival should either be assigned a set of items in which the fraction of items from each group is within specified bounds or be assigned no items; the goal is to assign items to platforms in order to maximize the number of items assigned to platforms. In the second problem, we study online matchings with diversity constraints, i.e. for each platform, absolute lower bounds are specified for each group. Each platform on arrival should either be assigned a set of items that satisfy these bounds or be assigned no items; the goal is to maximize the set of platforms that get matched. We study approximation algorithms and hardness results for these problems. The technical core of our proofs is a new connection between these problems and the problem of matchings in hypergraphs. Our experimental evaluation shows the performance of our algorithms on real-world and synthetic datasets exceeds our theoretical guarantees.

Backdoor attacks are a serious security risk for graph learning models. Backdoors can be embedded in the target model by inserting backdoor triggers into the training dataset, causing the model to make incorrect predictions when the trigger is present. To counter backdoor attacks, backdoor detection has been proposed. An emerging detection strategy in the vision and NLP domains is based on an intriguing phenomenon: when training models on a mixture of backdoor and clean samples, the loss on backdoor samples drops significantly faster than on clean samples, allowing backdoor samples to be easily detected by selecting samples with the lowest loss values. However, the ignorance of topological feature information on graph data limits its detection effectiveness when directly applied to the graph domain. To this end, we propose an explanation-guided backdoor detection method to take advantage of the topological information. Specifically, we train a helper model on the graph dataset, feed each graph sample into the model, and then adopt explanation methods to attribute model prediction to an important subgraph. We observe that backdoor samples have distinct attribution distribution than clean samples, so the explanatory subgraph could serve as more discriminative features for detecting backdoor samples. Comprehensive experiments on multiple popular datasets and attack methods demonstrate the effectiveness and explainability of our method. Our code is available at https://anonymous.4open.science/r/GNN_backdoor_detection-2BBF.

Value decomposition methodologies have gained considerable traction in the domain of multi-agent reinforcement learning. However, in environments characterized by sparse rewards, agents encounter difficulties in assessing the effectiveness of their actions in achieving the global objective, thereby impeding the algorithm's convergence rate and overall efficacy. To tackle this issue, we propose IPERS, an efficient algorithm that utilizes Individual Prioritized Experience Replay with Subgoals for Sparse Reward Multi-Agent Reinforcement Learning. IPERS incorporates stochastic episode selection from experience replay and employs Q-learning to identify individual subgoals for each agent. By leveraging intrinsic rewards, the proposed IPERS method guides agents towards these subgoals, mitigating the adverse consequences of sparse rewards. Additionally, IPERS decomposes the joint action and combines it with prioritized experience replay to facilitate each agent's learning of state transitions. Experimental evaluations of IPERS in the SMAC environment demonstrate its swift adaptability and significant improvements in convergence performance and win rate across diverse multi-agent tasks, surpassing state-of-the-art algorithms

In peer review systems, reviewers are often asked to evaluate various features of submissions, such as technical quality or novelty. A score is given to each of the predefined features and based on these the reviewer has to provide an overall quantitative recommendation. It may be assumed that each reviewer has her own mapping from the set of features to a recommendation, and that different reviewers have different mappings in mind. This introduces an element of arbitrariness known as commensuration bias. In this paper we discuss a framework, introduced by Noothigattu, Shah and Procaccia, and then applied by the organizers of the AAAI 2022 conference. Noothigattu, Shah and Procaccia proposed to aggregate reviewer's mapping by minimizing certain loss functions, and studied axiomatic properties of this approach, in the sense of social choice theory. We challenge several of the results and assumptions used in their work and report a number of negative results. On the one hand, we study a trade-off between some of the axioms proposed and the ability of the method to properly capture agreements of the majority of reviewers. On the other hand, we show that dropping a certain unrealistic assumption has dramatic effects, including causing the method to be discontinuous.

Reinforcement learning (RL) has been developed for mean field games over graphs ($\mathcal{G}$-MFG)  in social media and network economics, in which the transition of agents between a node pair incurs an instantaneous reward. However, agents' en-route choices on edges are largely neglected that incur an experienced reward depending on agents' actions and population evolution along edges. Here we focus on a broader class of MFGs, named "dual MFG on graphs" ($\mathcal{G}$-dMFG), which models two interacting MFGs, namely, one on edges and one at nodes over a graph. In this setting, agents select travel speed along edges and next-go-to edge at nodes for a minimum cumulative cost, which arises from the congestion effect when many agents compete for the same resource. This has various implications for autonomous driving navigation, spatial resource allocation, and internet packet routing. We establish formally that $\mathcal{G}$-dMFG is a generic $\mathcal{G}$-MFG, encompassing a more complex cost structure (that is nonseparable between states and actions) and with no need to pre-specify a termination time horizon. RL algorithms are designed to solve mean field equilibria (MFE) on large-sized networks.

This paper paves the way for interpretable and configurable semantic similarity search, by training state-of-the-art models for identifying textual similarity guided by a set of aspects or dimensions. The similarity models are analyzed as to which interpretable dimensions of similarity they place the most emphasis on. We conceptually introduce configurable similarity search for finding documents similar in specific aspects but dissimilar in others. To evaluate the interpretability of these dimensions, we experiment with downstream retrieval tasks using weighted combinations of these dimensions. Configurable similarity search is an invaluable tool for exploring datasets and will certainly be helpful in many applied natural language processing research applications.

Responsibility anticipation is the process of determining if the actions of an individual agent may cause it to be responsible for a particular outcome. This can be used in a multi-agent planning setting to allow agents to anticipate responsibility in the plans they consider. The planning setting in this paper includes partial information regarding the initial state and considers formulas in linear temporal logic as positive or negative outcomes to be attained or avoided. We firstly define attribution for notions of active, passive and contributive responsibility, and consider their epistemic variants. We then use these to define the notion of responsibility anticipation. We prove that our notions of anticipated responsibility can be used to coordinate agents in a planning setting and give complexity results for our model, including showing that anticipation and attribution for passive responsibility are equivalent to classical planning. We also present an outline for solving some of our attribution and anticipation problems using PDDL solvers.

We propose the use of controlled natural language as a target for knowledge graph question answering (KGQA) semantic parsing via language models as opposed to using formal query languages directly. Controlled natural languages are close to (human) natural languages, but can be unambiguously translated into a formal language such as SPARQL. Our research hypothesis is that the pre-training of large language models (LLMs) on vast amounts of textual data leads to the ability to parse into controlled natural language for KGQA with limited training data requirements. We devise an LLM-specific approach for semantic parsing to study this hypothesis. Our approach takes advantage of the hallucination and self-reflection capabilities of LLMs for relation linking. To conduct our study, we created a dataset that allows the comparison of one formal and two different controlled natural languages. Our analysis shows that training data requirements are indeed substantially reduced when using controlled natural languages, which is of paramount importance since collecting and maintaining high-quality KGQA semantic parsing training data is very expensive and time-consuming.

Answering temporal CQs over temporalized Description Logic knowledge bases (TKB) is a main technique to realize ontology-based situation recognition. In case the collected data in such a knowledge base is inaccurate, important query answers can be missed. In this paper we introduce the TKB Alignment problem, which computes a variant of the TKB that minimally changes the TKB, but entails the given temporal CQ and is in that sense (cost-)optimal. We investigate this problem for ALC TKBs and conjunctive queries with LTL operators and devise a solution technique to compute (cost-optimal) alignments of TKBs that extends techniques for the alignment problem for propositional LTL over finite traces.

Agent Scheduling Problems (ASPs) are common in various real-world situations, requiring explainable decision-making processes to effectively allocate resources to multiple agents while fostering understanding and trust. To address this need, this paper presents a logic-based framework for providing explainable decisions in ASPs. Specifically, the framework addresses two types of queries: reason-seeking queries, which explain the reasoning behind scheduling decisions, and modification-seeking queries, which offer guidance on making infeasible decisions feasible. Acknowledging the importance of privacy in multi-agent scheduling, we introduce a privacy-loss function that measures the disclosure of private information in explanations, enabling a privacy-preserving aspect in our framework. By using this function, we introduce the notion of privacy-aware explanations and present an algorithm for computing them. Empirical evaluations demonstrate the effectiveness and versatility of our approach.

Modelling causal responsibility in multi-agent spatial interactions is crucial for safety and efficiency of interactions of humans with autonomous agents. However, current formal metrics and models of responsibility either lack grounding in ethical and philosophical concepts of responsibility, or cannot be applied to spatial interactions. In this work we propose a metric of causal responsibility which is tailored to spatial interactions. In such interactions, a given agent can, by reducing another agent's feasible action space, influence the latter. Therefore, we propose feasible action space reduction (FeAR) as a metric for causal responsibility among agents. Specifically, we look at ex-post causal responsibility for simultaneous actions. We propose the use of moves de rigueur - a consistent set of prescribed actions for agents - to model the effect of norms on responsibility allocation. We applied the metric in a grid world simulation for spatial interactions and show how the actions, contexts and norms affect the causal responsibility ascribed to agents. We demonstrate the application of this metric in complex multi-agent interactions. We argue that the FeAR metric is a step towards an interdisciplinary framework for quantifying responsibility that is needed to ensure safety and meaningful human control in human-AI systems.

This paper focuses on graphical models for modelling preferences in combinatorial space and their use for item optimization. The preferential optimization task seeks to find the preferred item that contains some defined values, which is useful for many recommendation settings found in e-commerce. We show that efficient (i.e., with polynomial time complexity) preferential optimization is achieved with a subset of cyclic CP-nets called conditional acyclic CP-net. We also introduce a new graphical preference model, called Conditional-Optimality networks (CO-networks), that are more succinct than conditional acyclic CP-nets and LP-trees but have the same expressiveness with respect to optimization. Finally, we show that preferential optimization can be used for encoding alternatives into partial instantiations, and vice versa, paving the way towards CO-nets and CP-nets unsupervised learning with the minimal description length (MDL) principle.

Rejection sampling is a common tool, often touted as an ``easy'' way to obtain valid samples from a distribution $f(\cdot)$ of interest. In practice it is non-trivial to apply, often requiring considerable mathematical effort to devise a good proposal distribution $g(\cdot)$ and select a supremum $C$. More advanced samplers require additional mathematical derivations, limitations on $f(\cdot)$, or even cross-validation, making them difficult to apply. We devise a new approximate baseline approach to rejection sampling that works with less information, requiring only a differentiable $f(\cdot)$ be specified, making it easier to use. We propose a new approach to rejection sampling by refining a parameterized proposal distribution with a loss derived from the acceptance threshold. In this manner we obtain comparable or better acceptance rates on current benchmarks by up to $7.3\times$, while requiring no extra assumptions or any derivations to use: only a differentiable $f(\cdot)$ is required. While approximate, the results are correct with high probability, and in all tests pass a distributional check. This makes our approach easy to use, reproduce, and efficacious.

Envy-freeness is one of the prominent fairness notions in fair multiagent resource allocation but it has been mainly studied from an individual point of view. When the agents are partitioned into groups, fairness between groups is desirable. Several notions of group envy-freeness have been proposed over the last few years in the domain of fair division. In this paper we show that when groups may have different sizes and each agent gets at most one item, existing group envy-freeness notions fail to satisfy some desirable axioms. This motivates us to propose an original notion of degree of envy-freeness among groups, based on the counterfactual comparison of sub-groups of the same size. While this notion is computationally demanding, we show that it can be efficiently approximated thanks to an adapted sampling method, thus showing that our approach can be of practical relevance.

This paper introduces a novel pipeline for summarising timelines of events reported by multiple news sources. Transformerbased models for abstractive summarisation generate coherent and concise summaries of long documents but can fail to outperform established extractive methods on specialised tasks such as timeline summarisation (TLS). While extractive summaries are more faithful to their sources, they may be less readable and contain redundant or unnecessary information. This paper proposes a preference-based reinforcement learning (PBRL) method for adapting pretrained abstractive summarisers to TLS, which can overcome the drawbacks of extractive timeline summaries. We define a compound reward function that learns from keywords of interest and pairwise preference labels, which we use to fine-tune a pretrained abstractive summariser via offline reinforcement learning. We carry out both automated and human evaluation on three datasets, finding that our method outperforms a comparable extractive TLS method on two of the three benchmark datasets, and participants prefer our method’s summaries to those of both the extractive TLS method and the pretrained abstractive model. The method does not require expensive reference summaries and needs only a small number of preferences to align the generated summaries with human preferences.

Continual learning (CL) is a challenging machine learning setting that is attracting the interest of an increasing number of researchers. Among recent CL works, architectural strategies appear particularly promising due to their potential to expand and adapt the model architecture as new tasks are presented. However, existing solutions do not efficiently exploit model sparsity due to the adoption of constant pruning ratios. Moreover, current approaches exhibit a tendency to quickly saturate model capacity since the number of weights is limited and each weight is restricted to a single value. In this paper, we propose Ada-QPacknet, a novel architectural CL method that resorts to adaptive pruning and quantization. These two features allow our model to overcome the two crucial issues of effective exploitation of model sparsity and efficient use of model capacity. Specifically, adaptive pruning restores model capacity by reducing the number of weights assigned to each task to a smaller subset of weights that preserves the performance of the full set, allowing other weights to be used for future tasks. Adaptive quantization separates each weight into multiple components with adaptively reduced bit-width, allowing a single weight to solve more than one task without significant performance drops, leading to improved exploitation of model capacity. Experimental results on benchmark CL scenarios show that our proposed method achieves better results in terms of accuracy than existing rehearsal, regularization, and architectural CL strategies. Moreover, our method significantly outperforms forget-free competitors in terms of efficient exploitation of model capacity.

Physics-informed neural networks (PINNs) incorporate physical knowledge from the problem domain as a soft constraint on the loss function, but recent work has shown that this can lead to optimization difficulties. Here, we study the impact of the location of the collocation points on the trainability of these models. We find that the vanilla PINN performance can be significantly boosted by adapting the location of the collocation points as training proceeds. Specifically, we propose a novel adaptive collocation scheme which progressively allocates more collocation points (without increasing their number) to areas where the model is making higher errors (based on the gradient of the loss function in the domain). This, coupled with a judicious restarting of the training during any optimization stalls (by simply resampling the collocation points in order to adjust the loss landscape) leads to better estimates for the prediction error. We present results for several problems, including a 2D Poisson and diffusion-advection system with different forcing functions. We find that training vanilla PINNs for these problems can result in up to 70% prediction error in the solution, especially in the regime of low collocation points. In contrast, our adaptive schemes can achieve up to an order of magnitude smaller error, with similar computational complexity as the baseline. Furthermore, we find that the adaptive methods consistently perform on-par or slightly better than vanilla PINN method, even for large collocation point regimes. The code for all the experiments has been anonymously open-sourced.

Symbolic knowledge-extraction (SKE) techniques are becoming of key importance for AI applications since they enable the explanation of opaque black-box predictors, enhancing trust and transparency. Among all the available SKE techniques, the best option for the case at hand should be selected. However, an automatic comparison between different options can be performed only if an adequate metric -- such as a scoring function resuming all the interesting features of the extractors -- is provided. The definition of evaluation metrics for symbolic knowledge extractors is currently neglected in the literature. Accordingly, in this paper we introduce the FiRe score metric to assess the quality of a symbolic knowledge-extraction procedure, taking into account both its predictive performance and the readability of the extracted knowledge. It is compared to another existing scoring metric and a rigorous mathematical formulation is provided along with several practical examples to highlight its effectiveness to the end of being exploited inside automatic hyper-parameter tuning procedures.

Model-based diagnosis is a powerful, versatile and well-founded approach to troubleshooting a wealth of different types of systems. Diagnosis algorithms are both numerous and highly heterogeneous. In this work, we propose a taxonomy that allows their standardized assessment, classification and comparison. The aim is to (i) give researchers and practitioners an impression of the diverse landscape of available techniques, (ii) allow them to easily retrieve and compare the main features as well as pros and cons, and (iii) facilitate the selection of the “right” algorithm to adopt for a particular problem case, e.g., in practical diagnostic settings, for comparison in experimental evaluations, or for reuse, modification, extension, or improvement in the course of research. Finally, we demonstrate the value and application of the taxonomy by assessing and categorizing a range of more than 30 important diagnostic methods, and we point out how using the taxonomy as a common guideline for algorithm analysis would benefit the research community in various regards.

Centralized Multi-Agent Reinforcement Learning (MARL) presents itself as an ideal framework for aggregation companies (e.g., Uber, Lyft, Deliveroo) that have to take a sequential set of centralized decisions on assigning individual agents (typically resources like taxis, food delivery personnel) to customer requests online in the presence of demand uncertainty. However, centralized learning is especially challenging in such very large scale environments, with thousands of agents/resources and hundreds of thousands of requests coming in each day. In this paper, we provide a novel value decomposition mechanism that is able to tackle the scale and provide high quality (matching) decisions at each time step. We show that our value decomposition approach, Conditional Expectation
based Value Decomposition (CEVD) is more sustainable (requires 9.9% fewer vehicles to serve equal number of requests) and more efficient (serves 9.76% more requests, while traveling 13.32% lesser distance) than the current best approach over two different city scale (New York and Chicago) benchmarks for ride pooling using taxis.

(Gradient) Expectation Maximization (EM) is a widely used algorithm for estimating the maximum likelihood of mixture models or incomplete data problems. A major challenge facing this popular technique is how to effectively preserve the privacy of sensitive data. Previous research on this problem has already lead to the discovery of some Differentially Private (DP) algorithms for (Gradient) EM. However, unlike in the non-private case, existing techniques are not yet able to provide finite sample statistical guarantees. To address this issue, we propose in this paper the first DP version of Gradient EM algorithm with statistical guarantees. Specifically, we first propose a new mechanism for privately estimating the mean of a heavy-tailed distribution, which significantly improves a previous result in \cite{wangicml2020}, and it could be extended to the local DP model, which has not been studied before. Next, we apply our general framework to three canonical models: Gaussian Mixture Model (GMM), Mixture of Regressions Model (MRM) and Linear Regression with Missing Covariates (RMC). Specifically, for GMM in the DP model, our estimation error is near optimal in some cases. For the other two models, we provide the first result on finite sample statistical guarantees. Our theory is supported by thorough numerical experiments on both real-world data and synthetic data.

Complex Event Recognition (CER) systems detect event occurrences in streaming input using predefined event patterns. Techniques that learn event patterns from data highly desirable in CER. Since such patterns are typically represented by symbolic automata, we propose a family of such automata where the transition-enabling conditions are defined by Answer Set Programming (ASP) rules, and which, thanks to the strong connections of ASP to symbolic learning, are learnable from data. We present such a learning approach in ASP, capable of jointly learning the structure of an automaton, while synthesizing the transition guards from building-block predicates, and a scalable, incremental version thereof that progressively revises models learnt from mini-batches using Monte Carlo Tree Search. We evaluate our approach on three CER datasets and empirically demonstrate its efficacy.

Reinforcement learning (RL) is a powerful approach to training agents to perform tasks, but designing an appropriate reward mechanism is critical to its success. However, in many cases, the complexity of the learning objectives goes beyond the capabilities of the Markovian assumption, necessitating a more sophisticated reward mechanism. Reward machines and omega-regular languages are two formalisms used to express non-Markovian rewards for quantitative and qualitative objectives, respectively. This paper introduces omega-regular reward machines, which integrate reward machines with omega-regular languages to enable an expressive and effective reward mechanism for RL. We present a model-free RL algorithm to compute epsilon-optimal strategies against omega-regular reward machines and evaluate the effectiveness of the proposed algorithm through experiments.

The use of social media during election campaigns has become increasingly popular. However, the unbridled nature of online discourse can lead to the propagation of hate speech, which has far-reaching implications for the democratic process. Natural Language Processing (NLP) techniques are being used to counteract the spread of hate speech and promote healthy online discourse. Despite the increasing need for NLP techniques to tackle hate speech, research on low-resource languages like Nepali is limited, posing a challenge to the realization of the United Nations' Leave No One Behind principle, which calls for inclusive development that benefits all individuals and communities, regardless of their backgrounds or circumstances. To bridge this gap, we introduce NEHATE, a large-scale manually annotated dataset of hate speech and its targets in Nepali local election discourse. The dataset comprises 13,505 tweets, annotated for hate speech with further sub-categorization of hate speech into targets such as community, individual, and organization. Benchmarking of the dataset with various algorithms has shown potential for improvement in performance. We have made the dataset publicly available to promote further research and development, while also contributing to the UN SDGs aimed at fostering peaceful, inclusive societies and justice and strong institutions.

Algorithms to solve hard combinatorial problems often exhibit complementary performance, i.e. where one algorithm fails, another shines. Algorithm portfolios and algorithm selection take advantage of this by running all algorithms in parallel or choosing the best one to run on a problem instance. In this paper, we show that neither of these approaches gives the best possible performance and propose the happy medium of running a subset of all algorithms in parallel. We propose a method to choose this subset automatically for each problem instance, and demonstrate empirical improvements of up to 19% in terms of runtime, 81% in terms of misclassification penalty, and 26% in terms of penalized averaged runtime on scenarios from the ASlib benchmark library. Unlike all other algorithm selection and scheduling approaches in the literature, our performance measures are based on the actual performance for algorithms running in parallel rather than assuming overhead-free parallelization based on sequential performance. Our approach is easy to apply in practice and does not require to solve hard problems to obtain a schedule, unlike other techniques in the literature, while still delivering superior performance.

Supplying data augmentation to conversational question answering (CQA) can effectively improve model performance. However, there is less improvement from single-turn datasets in CQA due to the distribution gap between single-turn and multi-turn datasets. On the other hand, while numerous single-turn datasets are available, we have not utilized them effectively. To solve this problem, we propose a novel method to convert single-turn datasets to multi-turn datasets. The proposed method consists of three parts, namely, a QA pair Generator, a QA pair Reassembler, and a question Rewriter. Given a sample consisting of context and single-turn QA pairs, the Generator obtains candidate QA pairs and a knowledge graph based on the context. The Reassembler utilizes the knowledge graph to get sequential QA pairs, and the Rewriter rewrites questions from a conversational perspective to obtain a multi-turn dataset S2M. Our experiments show that our method can synthesize effective training resources for CQA. Notably, S2M ranks 1st place on the QuAC leaderboard1 at the time of submission (Aug 24th, 2022).

Judgment summaries are beneficial for legal practitioners to comprehend and retrieve case law efficiently. Unlike summaries in general domains, e.g., news, judgment summaries often requires a clear structure. Such a structure helps readers grasp the information contained in the summary and reduces information loss. To the best of our knowledge, none of the existing text summarizers can generate summaries aligned with the summary structure in the legal domain. Inspired by this observation, this paper introduces a Summary Structure-Enhanced (SSE) method to synthesize structured summaries for legal documents. SSE can easily be incorporated into the Encoder-Decoder framework, which is commonly adopted in state-of-the-art text summarizers. Experiments on the datasets of New Zealand and Chinese judgments show that the proposed method consistently improves the performance of state-of-the-art summarizers in terms of Rouge scores.

We consider an auction design problem where a seller sells multiple homogeneous items to a set of connected buyers. Each buyer only knows the buyers she directly connects with and has a diminishing marginal utility valuation for the items. The seller initially only connects to some buyers who can be directly invited to the sale by the seller. Our goal is to design an auction to incentivize the buyers who are aware of the auction to further invite their neighbors to join the auction. This is challenging because the buyers are competing for the items and they would not invite each other by default. Thus, rewards need to be given to buyers who diffuse information, but the rewards should be carefully designed to guarantee both invitation incentives and the seller's revenue. Solutions have been proposed recently for the settings where each buyer requires at most one unit and demonstrated the difficulties of the design. We move this forward to propose the very first diffusion auction for the multi-unit demand settings to improve both the social welfare and the seller's revenue.

Offline reinforcement learning (RL) circumvents costly interactions with the environment by utilising historical trajectories. Incorporating a world model into this method could substantially enhance the transfer performance of various tasks without expensive calculations from scratch. However, due to the complexity arising from different types of generalisation, previous works have focused almost exclusively on single-environment tasks. In this study, we introduce a multi-environment offline RL setting to investigate whether a generalised world model can be learned from large, diverse datasets and serve as a good surrogate for policy learning in different tasks. Inspired by the success of multi-task prompt methods, we propose the Task-prompt Generalised World Model (TGW) framework, which demonstrates notable performance in this setting. TGW comprises three modules: a task-state prompter, a generalised dynamics module, and a reward module. We implement the generalised dynamics module as a transformer-based recurrent state-space model (TransRSSM) and employ prompts to provide task-specific instructions, enabling TGW to address the internal stochasticity of the generalised world model. On the MuJoCo control benchmarks, TGW significantly outperforms previous offline RL algorithms in multi-environment setting. 

Automatic radiology report generation is challenging as medical images or reports are usually similar to each other due to the common content of anatomy. This makes a model hard to capture the uniqueness of individual images and is prone to producing undesired generic or mismatched reports. This situation calls for learning more discriminative features that could capture even fine-grained mismatches between images and reports. To achieve this, this paper proposes a novel framework to learn discriminative image and report features by distinguishing them from their closest peers, i.e., hard negatives. Especially, to attain more discriminative features, we gradually raise the difficulty of such a learning task by creating increasingly hard negative reports for each image in the feature space during training, respectively. By treating the increasingly hard negatives as auxiliary variables, we formulate this process as a min-max alternating optimisation problem. At each iteration, conditioned on a given set of hard negative reports, image and report features are learned as usual by minimising the loss functions related to report generation. After that, a new set of harder negative reports will be created by maximising a loss reflecting image-report alignment. By solving this optimisation, we attain a model that can generate more specific and accurate reports. It is noteworthy that our framework enhances discriminative feature learning without introducing extra network weights.   Also, in contrast to the existing way of generating hard negatives, our framework extends beyond the granularity of the dataset by generating harder samples out of the training set. Experimental study on benchmark datasets verifies the efficacy of our framework and shows that it can serve as a plug-in to readily improve existing medical report generation models.

Federated Learning allows collaborative training without data sharing in settings where participants do not trust the central server and one another. Privacy can be further improved by ensuring that communication between the participants and the server is anonymized through a shuffle; decoupling the participant id from their data. This paper seeks to examine whether such a defense is adequate to guarantee anonymity, by proposing a novel fingerprinting attack over gradients sent by the participants to the server. We show that clustering of gradients can easily break the anonymization in an empirical study of learning federated language models on two language corpora. We then show that training with differential privacy can provide a practical defense against our fingerprint attack.

Earley Parser is a top-down parser proposed for context-free grammars and used, for example, in the grammar constraint. Parsing trees of context-free grammars are very close to task decomposition trees used in hierarchical planning, specifically when the actions are totally ordered. This paper suggests using the Earley Parser to recognize totally ordered hierarchical plans. Given a sequence of actions - a prefix of the plan - and a task decomposition model, the plan recognition problem asks which task decomposes to a plan containing the given plan prefix. We will show that the Earley parser significantly increases the speed of plan recognition compared to the existing bottom-up parsing-based plan recognizer. The Earley parser's performance is also on a par with the planning-based plan recognizer despite not using any planning heuristics.

In the real world, the execution of the actions planned for an agent is never guaranteed to succeed, as they can fail in a number of unexpected ways that are not explicitly captured in the planning model. Based on these observations, we introduce the task of finding plans for classical planning that are resilient to action execution failures. We refer to this problem as Resilient Planning and to its solutions as K-resilient plans; such plans guarantee that an agent will always be able to reach its goals (possibly by replanning alternative sequences of actions) as long as no more than K failures occur along the way. We also present ResPlan, a new algorithm for Resilient Planning, and we compare its performance to methods based on compiling Resilient Planning to FOND planning.

In neural architecture search (NAS), training every sampled architecture is very time-consuming and should be avoided. Weight-sharing is a promising solution to speed up the evaluation process. However, training the supernetwork incurs many discrepancies between the actual ranking and the predicted one. Additionally, efficient deep-learning engineering processes require incorporating realistic hardware-performance metrics into the NAS evaluation process, also known as hardware-aware NAS (HW-NAS). In HW-NAS, estimating task-specific performance and hardware efficiency are both required. This paper proposes a supernetwork training methodology that preserves the Pareto ranking between its different subnetworks resulting in more efficient and accurate neural networks for a variety of hardware platforms. The results show a 97% near Pareto front approximation in less than 2 GPU days of search, which provides 2x speed up compared to state-of-the-art methods. We validate our methodology on NAS-Bench-201, DARTS, and ImageNet. Our optimal model achieves 77.2% accuracy (+1.7% compared to baseline) with an inference time of 3.68ms on Edge GPU for ImageNet, which yields a 2.3x speedup.  Training implementation can be found: https://anonymous.4open.science/r/PRP-NAS-7CA6/

Extracting image semantics effectively and assigning corresponding labels to multiple objects or attributes for natural images is challenging due to the complex scene contents and confusing label dependencies. Recent works have focused on modeling label relationships with graph and understanding object regions using class activation maps (CAM). However, these methods ignore the complex intra- and inter-category relationships among specific semantic features, and CAM is prone to generate noisy information. To this end, we propose a novel semantic-aware dual contrastive learning framework that incorporates sample-to-sample contrastive learning (SSCL) as well as prototype-to-sample contrastive learning (PSCL). Specifically, we leverage semantic-aware representation learning to extract category-related local discriminative features and construct category prototypes. Then based on SSCL, label-level visual representations of the same category are aggregated together, and features belonging to distinct categories are separated. Meanwhile, we construct a novel PSCL module to narrow the distance between positive samples and category prototypes and push negative samples away from the corresponding category prototypes. Finally, the discriminative label-level features related to the image content are accurately captured by the joint training of the above three parts. Experiments on five challenging large-scale public datasets demonstrate that our proposed method is effective and outperforms the state-of-the-art methods.  Code and supplementary materials are released on https://github.com/yu-gi-oh-leilei/SADCL.

Coordination and joint ability are important problems in representation and reasoning about multi-agent systems. Ghaderi et al. presented a formalization of joint ability of coalitions in the expressive first-order language of the situation calculus. Essentially, a coalition has joint ability to achieve a goal if after iterated elimination of dominated strategies, any remaining joint strategy achieves the goal. Based on their work, Liu et al. proposed JAADL, a modal logic for joint abilities under strategy commitments. In this paper, we propose EJAADL, an epistemic extension of JAADL, for imperfect information games where agents may have incomplete knowledge or even false beliefs about the world. Like Ghaderi et al.’s work, elimination of dominated strategies is now based on beliefs about the world, rather than facts about the world as in JAADL. Strategies are required to be uniform, i.e., they select the same action in all accessible histories. We illustrate EJAADL with examples, analyze its properties, and show that model checking memoryless EJAADL is in EXPTIME. Morover, we consider the fragment of memoryless EJAADL with only bounded number of elimination of dominated strategies, and show that the model checking problem for this fragment can be done in PSPACE.

The orthogonality constraints, including the hard and soft ones, have been used to normalize the weight matrices of Deep Neural Network (DNN) models, especially the Convolutional Neural Network (CNN) and Vision Transformer (ViT), to reduce model parameter redundancy and improve training stability. However, the robustness to noisy data of these models with constraints is not always satisfactory. In this work, we propose a novel two-stage approximately orthogonal training framework (TAOTF) to find a trade-off between the orthogonal solution space and the main task solution space to solve this problem in noisy data scenarios. In the first stage, we propose a novel algorithm called polar decomposition-based orthogonal initialization (PDOI) to find a good initialization for the orthogonal optimization. In the second stage, unlike other existing methods, we apply soft orthogonal constraints for all layers of DNN model. We evaluate the proposed model-agnostic framework both on the natural image and medical image datasets, which show that our method achieves stable and superior performances to existing methods.

Knowledge-grounded dialogue (KGD) has become increasingly essential for online services, enabling individuals to obtain desired information. While KGD contains knowledge information, most knowledge points are fragmented and repeated in dialogues, making it difficult for users to quickly grasp complete and key information from a collection of sessions. In this paper, we propose a novel task of dialogue-grounded knowledge points generation (DialKPG) to condense a collection of sessions on a topic into succinct and complete knowledge points. To enable empirical study, we create TopicDial and OpenDial corpus based on two existing knowledge-grounded dialogue corpus FaithDial and OpenDialKG by a Three-Stage Annotation Framework, and establish a novel approach for DialKPG task, namely MSAM (Multi-Level Salience-Aware Mixture). MSAM explicitly incorporates salient information at the token-level, utterance-level, and session-level to better guide knowledge points generation. Extensive experiments have verified the effectiveness of our method over competitive baselines. Furthermore, our analysis shows that the proposed model is particularly effective at handling long inputs and multiple sessions due to its strong capability of duplicated elimination and knowledge integration.

The vulnerability of artificial intelligence has become a bottleneck, with adversarial attacks posing a significant threat to natural language processing. Although multiple defense mechanisms have been proposed, they often suffer from strong constraints, weak generalization, and low scalability. To address these issues, we propose leveraging perplexity to quantify the difference between clean and adversarial examples based on the observation of numerous cases. We then statistically prove the significant difference between them using Bayesian hypothesis testing.
Subsequently, we develop an adversarial defense framework named UMPS, which contains two branches: “Uncovering the Mask”(UM) and “Perplexity-guided Sampling”(PS). UM utilizes a masked language model and JW distance to recover out-of-vocabulary words, while PS employs perplexity to locate the optimal sample within a convex hull which is constructed with integrated gradients. Theoretically, the proposed framework fulfills three requirements: effectiveness, universality, and portability. The experimental results demonstrate that UMPS effectively enhances the robustness of language models including Bert, against advanced attacks and outperforms three baselines. Furthermore, the ablation study is conducted to support the validity and necessity of the two branches, and the post-hoc test of the difference in perplexity explains the defense performance of our framework.

Motivated by the abundance of biomedical publications and the need to better understand the relationship between food and health, we study a new sentiment analysis task based on literature- based discovery. Many attempts have been made to introduce health into recipe recommendation systems and food analysis. However, these methods focus highly on ingredient nutritional components or use simple computational models trained on curated labeled data. With the high availability of food and health information in biomedical texts but with a high cost of data labeling, there is a need for enhanced models to capture the intrinsic relationship between food ingredients and biomedical concepts, including genes, diseases, nutrition, and chemicals. This model should also utilize both available labeled resources and unlabelled data. In this paper, we propose a new sentiment analysis task called Entity Relationship Sentiment Analysis (ERSA) that focuses on capturing the relationship between biomedical and food concepts. This task extends the popularly studied Aspect Based Sentiment Analysis (ABSA) task. Specifically, our study focuses on the sentiment analysis of (entity-entity) pairs given a biomedical text sentence. This task presents a more significant challenge than traditional sentiment analysis tasks, as the sentiment expressed in the sentence may not necessarily reflect the sentiment of the relationship between two given entities in the sentence. Furthermore, we propose a semi-supervised architecture that combines two different types of word embeddings to encode the ERSA task better. The experimental results demonstrate that our proposed method consistently outperforms other semi-supervised learning methods across various learning scenarios.

Belief change studies how an agent modifies her beliefs on receiving new information. However, so far most research on belief change works on beliefs represented in propositional logic. There have been many works on integrating belief revision with reasoning about actions, and some works extending belief change from propositional logic to epistemic logics. In this paper, we study revision on beliefs of a third person represented with the multi-agent KD45 logic. Our formal technique is analogous to that of distance-based belief revision in propositional logic: to revise a KB by a formula, select from models of the formula those that are closest to models of the KB. To this end, a challenge is that in modal logics, a formula may have infinitely many Kripke models. To tackle this, we propose a variant of Moss' canonical formulas called alternating canonical formulas, treat them as models for formulas, and define a notion of distance between them, based on the Hausdorff distance between two sets. We show that our revision satisfies all of the AGM postulates. To give syntactic characterizations of our revision, we make use of a normal form for KD45n called alternating cover disjunctive formulas (ACDFs). We give syntactic characterizations firstly on fragments of ACDFs called proper ACDFs and alternating cover conjunctive formulas (ACCFs), and finally on the whole ACDFs.

Using insights from parametric integer linear programming, we significantly improve a previous result [Proc. ACM EC 2019] on high-multiplicity fair allocation. Therein (answering an open question from a preceding work [Proc. IJCAI 2016]) the authors proved that the problem of finding envy-free Pareto-efficient allocations of indivisible items is fixed-parameter tractable with respect to the combined parameter “number of agents” plus “number of item types.” Our central improvement, compared to this result, is to break the condition that the corresponding utility and multiplicity values have to be encoded in unary required there. Concretely, we show that, while preserving fixed-parameter tractability, these values can be encoded in binary, thus greatly expanding the range of feasible values.

Advancements in Conversational Natural Language Processing (NLP) have the potential to address critical social challenges, particularly in achieving the United Nations' Sustainable Development Goal of quality education through the development of Virtual Tutors. However, the application of NLP in the educational domain, especially language learning, has been limited due to the inherent complexities of the field and the scarcity of available datasets. In this paper, we introduce T-VAKS (Tutoring Virtual Agent with Knowledge Selection), a novel language tutoring multimodal Virtual Agent (VA) designed to assist students in learning a new language. Our approach employs an information theory-based knowledge selection module built on top of a multimodal seq2seq generative model, facilitating the generation of appropriate, informative, and contextually relevant tutor responses. T-VAKS aims to bridge the gap between NLP and the educational domain, enabling more effective language tutoring through intelligent virtual agents. The knowledge selection module consists of two sub-modules: (i) knowledge relevance estimation, and (ii) knowledge focusing framework. We evaluate the performance of our proposed model against various baseline models and the most recent state-of-the-art models, using multiple evaluation metrics. The results demonstrate that T-VAKS outperforms competing models, highlighting the potential of our approach in enhancing language learning through the use of conversational NLP and virtual agents, ultimately addressing social challenges and promoting well-being.

Deep neural networks are susceptible to adversarial attacks, which pose a significant threat to their security and reliability in real-world applications. The most notable adversarial attacks are transfer-based attacks, where an adversary crafts an adversarial example to fool one model, which can also fool other models. While previous research has made progress in improving the transferability of untargeted adversarial examples, the generation of targeted adversarial examples that can transfer between models remains a challenging task. In this work, we present a novel approach to generate transferable targeted adversarial examples by exploiting the vulnerability of deep neural networks to perturbations on high-frequency components of images. We observe that replacing the high-frequency component of an image with that of another image can mislead deep models, motivating us to craft perturbations containing high-frequency information to achieve targeted attacks. To this end, we propose a method called Low-Frequency Adversarial Attack (LFAA), which trains a conditional generator to generate targeted adversarial perturbations that are then added to the low-frequency component of the image. Extensive experiments on ImageNet demonstrate that our proposed approach significantly outperforms state-of-the-art methods, improving targeted attack success rates by a margin from 3.2% to 15.5%.

Crowd flow prediction is vital in various fields such as traffic management, public safety, and urban planning. The main challenge in crowd flow prediction lies in effectively modeling the periodic temporal dependency and long-range spatial dependency. In the temporal domain, crowd flow shows a strong periodicity which is exploited by existing works to build multi-time-scale spatial-temporal features. However, these works hardly consider the disturbance of periods, that is, the crowd flow is not strictly periodic. In the spatial domain, existing works mainly utilize CNN to capture spatial dependency, but the small receptive field of the convolution operator limits the ability to capture the long-range dependency between crowd flows in different regions. In this paper, we propose GridFormer, a Transformer network, in which a periodically shifted sampling method and attention mechanism are employed to handle the temporal shifting in the daily and weekly periodicity, and a pyramid 3D Swin Transformers network is designed to capture long-range spatial dependency in a hierarchical manner. Meanwhile, the pyramid 3D Swin Transformers network jointly models spatial-temporal features to enable better interaction between the spatial and temporal domains. Experimental results on three crowd flow datasets demonstrate that our GridFormer outperforms the state-of-the-art crowd flow prediction methods.

Learning fair policies in reinforcement learning (RL) is important when the RL agent may impact many users. We investigate a variant of this problem where equity is still desired, but some users may be entitled to preferential treatment. In this paper, we formalize this more sophisticated fair optimization problem in deep RL using generalized fair social welfare functions (SWF), provide a theoretical discussion to justify our approach, explain how deep RL algorithms can be adapted to tackle it, and empirically validate our propositions on several domains. Our contributions are both theoretical and algorithmic, notably: (1) We obtain a general bound on the suboptimality gap in terms of SWF-optimality using average reward of a policy SWF-optimal for the discounted reward, which notably justifies using standard deep RL algorithms, even for the average reward; (2) Our algorithmic innovations  include a state-augmented DQN-based method for learning either deterministic or stochastic policies, which also applies to the usual fair optimization setting without any preferential treatment.

Medical Question Answering (MedQA) is one of the most popular and significant tasks in developing healthcare assistants. When humans extract an answer to a question from a document, they first (a) understand the question itself in detail and (b) utilize relevant knowledge/experiences to determine the answer segments. In multi-span question answering, it becomes increasingly important to comprehend the query accurately and possess relevant knowledge, as the interrelationship among different answer segments is essential for achieving completeness. Motivated by this, we first propose a transformer-based query semantic and knowledge ({\em QueSemKnow}) guided multi-span question-answering model. The proposed {\em QueSemKnow} works in a two-phased manner; in the first stage, a multi-task model is proposed to extract query semantics: (i) intent identification and (ii) question type prediction. In the second stage, {\em QueSemKnow} selects a relevant subset of the knowledge graph as the underlying context/document and extracts answers depending on the semantic information extracted from the first stage and context. We build a multi-task query semantic extraction model for query intent and query type identification to investigate the co-relation among these tasks. We also develop a semantically aware medical question-answering corpus called {\em QueSeMSpan MedQA}, where each question is tagged with its semantic information. The proposed model outperforms several baselines and existing state-of-the-art models by a large margin on multiple datasets, which firmly demonstrates the effectiveness of the human-inspired multi-span question-answering methodology. The dataset and code are available at https://github.com/AnonymousRW/QueSemKnow/.

In recent years, there has been growing interest in the application of temporal reasoning approaches and non-monotonic logics from artificial intelligence in dynamic systems that generate data. A well-known approach to temporal reasoning is the use of the progression technique, which allows for the online computation of logical consequences of a logical knowledge base over time. We consider the progression technique for Temporal Here and There and Temporal Equilibrium Logic, which is the logic underlying answer programming over linear-temporal logic (LTL). Compared to usual LTL online computation, where the goal is to check whether a trace is compliant with a temporal specification, our approach provides also the means to compute non-monotonic temporal reasoning over a trace of observations. Besides formal notions and results, we also present an algorithm for performing progression to monitor a dynamic system, which has been implemented as a proof of concept and allows for handling expressive application scenarios.

Strong and weak supervisions have complementary characteristics. However, utilizing both supervisions for named entity recognition (NER) has not been extensively studied. Moreover, the existing works address only incomplete annotations and neglects inaccurate annotations during NER model training. To effectively utilize weak labels, we introduce an auxiliary classifier that learns from weak labels. Furthermore, we adopt the teacher-student framework to handle both incomplete and inaccurate weak labels. A teacher model is first trained using both strongly and weakly supervised data, and next generates pseudo labels to replace weak labels. Then, the student model is trained so that the main classifier learns from both strong labels and confident pseudo labels while the auxiliary classifier learns from less confident pseudo labels. We also incorporate data augmentation through ChatGPT to generate additional annotated sentences to improve model performance and generalization capabilities. The experimental results with different weak supervisions demonstrate that our proposed method surpasses existing techniques.

Gradual semantics are methods that evaluate overall strengths of individual arguments in graphs. In this paper, we investigate gradual semantics for extended frameworks in which probabilities are used to quantify the uncertainty about arguments and attacks belonging to the graph. We define the likelihoods of an argument's possible strengths when facing uncertainty about the topology of the argumentation framework. We also define an approach to compare the strengths of arguments in this probabilistic setting. Finally, we propose a method to calculate the overall strength of each argument in the framework, and we evaluate this method against a set of principles.

In multi-task learning (MTL), multiple prediction tasks are learned jointly, such that generalization performance is improved by transferring information across the tasks. However, not all tasks are related, and training unrelated tasks together can worsen the prediction performance because of the phenomenon of negative transfer. To overcome this problem, we propose a novel MTL method that can robustly group correlated tasks into clusters and allow useful information to be transferred only within clusters. The proposed method is based on the assumption that the task clusters lie in the low-rank subspaces of the parameter space, and the number of them and their dimensions are both unknown. By applying subspace clustering to task parameters, parameter learning and task grouping can be done in a unified framework. To relieve the error induced by the basic linear learner and robustify the model, the effect of hidden tasks is exploited. Moreover, the framework is extended to a multi-layer architecture so as to progressively 
extract hierarchical subspace structures of tasks, which helps to further improve generalization. The optimization algorithm is proposed, and its effectiveness is validated by experimental results on both synthetic and real-world datasets.

In some applications, planning-monitoring systems generate plans and monitor their execution by other agents. During execution, agents might deviate from these plans for various reasons. The deviation from the expected behavior will be observed by the planning-monitoring system, which will replan in order to provide the agent a new suggested plan. Most existing replanning approaches maintain the goals and compute a plan that achieves them under
the new circumstances. This is often not realistic, as achieving the original goal might be very costly or impossible under the new circumstances. Furthermore, replanning approaches usually overlook agent’s behavior up to the observed deviation from the original plan. In this paper we introduce GREPLAN, a novel approach that proposes new replanning goals (and plans) by solving a multi-objective optimization problem that considers all goals within a perimeter of the original goal. Empirical results in several planning benchmarks show that GREPLAN successfully reacts to deviations from the original plan by generating new appropriate replanning goals.

Answer Set Programming (ASP) is a well-known AI formalism. Traditional ASP systems, that follow the “ground&Solve” approach, are intrinsically limited by the so-called grounding bottleneck. Basically, the grounding step (i.e., variable-elimination) can be computationally expensive and even unfeasible in several cases of practical interest. Recent work demonstrated that the grounding bottleneck can be partially overcome by compiling in external propagators subprograms acting as constraints. In this paper a novel compilation technique is presented that can be applied to tight normal programs; thus, the class of ASP programs that can be compiled is extended beyond constraints. The approach is implemented in the new system ProASP. ProASP skips entirely the grounding phase, and performs solving by injecting custom propagators in GLUCOSE. An experiment conducted on ground-intensive ASP benchmarks shows that ProASP is capable of solving instances that are out of reach for state-of-the-art ASP systems

Steganography, or hiding messages in plain sight, is a form of information hiding that is most commonly used for covert communication. As modern steganographic mediums include images, text, audio, and video, this communication method is being increasingly used by bad actors to propagate malware, exfiltrate data, and discreetly communicate. Current protection mechanisms rely upon steganalysis, or the detection of steganography, but these approaches are dependent upon prior knowledge, such as steganographic signatures from publicly available tools and statistical knowledge about known hiding methods. These dependencies render steganalysis useless against new or unique hiding methods, which are becoming increasingly common with the application of deep learning models. To mitigate the shortcomings of steganalysis, this work focuses on a deep learning sanitization technique called SUDS that is not reliant upon knowledge of steganographic hiding techniques and is able to sanitize universal and dependent steganography. SUDS is tested using least significant bit method (LSB), dependent deep hiding (DDH), and universal deep hiding (UDH). We demonstrate the capabilities and limitations of SUDS by answering five research questions, including baseline comparisons and an ablation study. Additionally, we apply SUDS to a real-world scenario, where it is able to increase the resistance of a poisoned classifier against attacks by 1375%.