AAAI 2024 论文列表

To achieve the ambitious goals of artificial intelligence, reinforcement learning must include planning with a model of the world that is abstract in state and time. Deep learning has made progress with state abstraction, but temporal abstraction has rarely been used, despite extensively developed theory based on the options framework. One reason for this is that the space of possible options is immense, and the methods previously proposed for option discovery do not take into account how the option models will be used in planning. Options are typically discovered by posing subsidiary tasks, such as reaching a bottleneck state or maximizing the cumulative sum of a sensory signal other than reward. Each subtask is solved to produce an option, and then a model of the option is learned and made available to the planning process. In most previous work, the subtasks ignore the reward on the original problem, whereas we propose subtasks that use the original reward plus a bonus based on a feature of the state at the time the option terminates. We show that option models obtained from such reward-respecting subtasks are much more likely to be useful in planning than eigenoptions, shortest path options based on bottleneck states, or reward-respecting options generated by the option-critic. Reward respecting subtasks strongly constrain the space of options and thereby also provide a partial solution to the problem of option discovery. Finally, we show how values, policies, options, and models can all be learned online and off-policy using standard algorithms and general value functions.

Constrained Reinforcement Learning employs trajectory-based cost constraints (such as expected cost, Value at Risk, or Conditional VaR cost) to compute safe policies. The challenge lies in handling these constraints effectively while optimizing expected reward. Existing methods convert such trajectory-based constraints into local cost constraints, but they rely on cost estimates, leading to either aggressive or conservative solutions with regards to cost. We propose an unconstrained formulation that employs reward penalties over states augmented with costs to compute safe policies. Unlike standard primal-dual methods, our approach penalizes only infeasible trajectories through state augmentation. This ensures that increasing the penalty parameter always guarantees a feasible policy, a feature lacking in primal-dual methods. Our approach exhibits strong empirical performance and theoretical properties, offering a fresh paradigm for solving complex Constrained RL problems, including rich constraints like expected cost, Value at Risk, and Conditional Value at Risk. Our experimental results demonstrate superior performance compared to leading approaches across various constraint types on multiple benchmark problems.

Reinforcement Learning (RL) has achieved remarkable success in safety-critical areas, but it can be weakened by adversarial attacks. Recent studies have introduced ``smoothed policies" to enhance its robustness. Yet, it is still challenging to establish a provable guarantee to certify the bound of its total reward. Prior methods relied primarily on computing bounds using Lipschitz continuity or calculating the probability of cumulative reward being above specific thresholds. However, these techniques are only suited for continuous perturbations on the RL agent's observations and are restricted to perturbations bounded by the l2-norm. To address these limitations, this paper proposes a general black-box certification method, called ReCePS, which is capable of directly certifying the cumulative reward of the smoothed policy under various lp-norm bounded perturbations. Furthermore, we extend our methodology to certify perturbations on action spaces. Our approach leverages f-divergence to measure the distinction between the original distribution and the perturbed distribution, subsequently determining the certification bound by solving a convex optimisation problem. We provide a comprehensive theoretical analysis and run experiments in multiple environments. Our results show that our method not only improves the tightness of certified lower bound of the mean cumulative reward but also demonstrates better efficiency than state-of-the-art methods.

This article considers the problem of diagnosing certain common errors in reward design. Its insights are also applicable to the design of cost functions and performance metrics more generally. To diagnose common errors, we develop 8 simple sanity checks for identifying flaws in reward functions. We survey research that is published in top-tier venues and focuses on reinforcement learning (RL) for autonomous driving (AD). Specifically, we closely examine the reported reward function in each publication and present these reward functions in a complete and standardized format in the appendix. Wherever we have sufficient information, we apply the 8 sanity checks to each surveyed reward function, revealing near-universal flaws in reward design for AD that might also exist pervasively across reward design for other tasks. Lastly, we explore promising directions that may aid the design of reward functions for AD in subsequent research, following a process of inquiry that can be adapted to other domains.

This proposal introduces an innovative AI-powered learning system designed to address educational disparities worldwide. Focused on developing countries, the system seamlessly translates educational content between English and native languages, breaking down language barriers. Leveraging advanced natural language processing and machine learning techniques, including transformer models like BERT and GPT-3, the system ensures inclusivity, effectiveness, and engagement. Built on prior research demonstrating AI's efficacy in language translation and personalized learning, the proposed system draws inspiration from successful projects like Duolingo Language Incubator. By providing inclusive and accessible learning experiences, it empowers individuals to overcome language barriers, fostering global participation. The potential impact is significant, with the system poised to accelerate learning, enhance literacy rates, and create a more skilled workforce in developing countries. This research reflects a commitment to revolutionize education through technology, aiming for lasting and transformative contributions to global society. Through AI-driven education, a brighter, more inclusive future is envisioned.

The Bahnar, a minority ethnic group in Vietnam with ancient roots, hold a language of deep cultural and historical significance. The government is prioritizing the preservation and dissemination of Bahnar language through online availability and cross-generational communication. Recent AI advances, including Neural Machine Translation (NMT), have transformed translation with improved accuracy and fluency, fostering language revitalization through learning, communication, and documentation. In particular, NMT enhances accessibility for Bahnar language speakers, making information and content more available. However, translating Vietnamese to Bahnar language faces practical hurdles due to resource limitations, particularly in the case of Bahnar language as an extremely low-resource language. These challenges encompass data scarcity, vocabulary constraints, and a lack of fine-tuning data. To address these, we propose transfer learning from selected pre-trained models to optimize translation quality and computational efficiency, capitalizing on linguistic similarities between Vietnamese and Bahnar language. Concurrently, we apply tailored augmentation strategies to adapt machine translation for the Vietnamese-Bahnar language context. Our approach is validated through superior results on bilingual Vietnamese-Bahnar language datasets when compared to baseline models. By tackling translation challenges, we help revitalize Bahnar language, ensuring information flows freely and the language thrives.

To trace the copyright of deep neural networks, an owner can embed its identity information into its model as a watermark. The capacity of the watermark quantify the maximal volume of information that can be verified from the watermarked model. Current studies on capacity focus on the ownership verification accuracy under ordinary removal attacks and fail to capture the relationship between robustness and fidelity. This paper studies the capacity of deep neural network watermarks from an information theoretical perspective. We propose a new definition of deep neural network watermark capacity analogous to channel capacity, analyze its properties, and design an algorithm that yields a tight estimation of its upper bound under adversarial overwriting. We also propose a universal non-invasive method to secure the transmission of the identity message beyond capacity by multiple rounds of ownership verification. Our observations provide evidence for neural network owners and defenders that are curious about the tradeoff between the integrity of their ownership and the performance degradation of their products.

In this paper, we focus on the open-set panoptic segmentation (OPS) task to circumvent the data explosion problem. Different from the close-set setting, OPS targets to detect both known and unknown categories, where the latter is not annotated during training. Different from existing work that only selects a few common categories as unknown ones, we move forward to the real-world scenario by considering the various tail categories (~1k). To this end, we first build a new dataset with long-tail distribution for the OPS task. Based on this dataset, we additionally add a new class type for unknown classes and re-define the training annotations to make the OPS definition more complete and reasonable. Moreover, we analyze the influence of several significant factors in the OPS task and explore the upper bound of performance on unknown classes with different settings. Furthermore, based on the analyses, we design an effective two-phase framework for the OPS task, including thing-agnostic map generation and unknown segment mining. We further adopt semi-supervised learning to improve the OPS performance. Experimental results on different datasets validate the effectiveness of our method.

Graph-based fraud detection (GFD) can be regarded as a challenging semi-supervised node binary classification task. In recent years, Graph Neural Networks (GNN) have been widely applied to GFD, characterizing the anomalous possibility of a node by aggregating neighbor information. However, fraud graphs are inherently heterophilic, thus most of GNNs perform poorly due to their assumption of homophily. In addition, due to the existence of heterophily and class imbalance problem, the existing models do not fully utilize the precious node label information. To address the above issues, this paper proposes a semi-supervised GNN-based fraud detector SEC-GFD. This detector includes a hybrid filtering module and a local environmental constraint module, the two modules are utilized to solve heterophily and label utilization problem respectively. The first module starts from the perspective of the spectral domain, and solves the heterophily problem to a certain extent. Specifically, it divides the spectrum into various mixed-frequency bands based on the correlation between spectrum energy distribution and heterophily. Then in order to make full use of the node label information, a local environmental constraint module is adaptively designed. The comprehensive experimental results on four real-world fraud detection datasets denote that SEC-GFD outperforms other competitive graph-based fraud detectors. We release our code at https://github.com/Sunxkissed/SEC-GFD.

Collaborative learning (CL) is a distributed learning framework that aims to protect user privacy by allowing users to jointly train a model by sharing their gradient updates only. However, gradient inversion attacks (GIAs), which recover users' training data from shared gradients, impose severe privacy threats to CL. Existing defense methods adopt different techniques, e.g., differential privacy, cryptography, and perturbation defenses, to defend against the GIAs. Nevertheless, all current defense methods suffer from a poor trade-off between privacy, utility, and efficiency. To mitigate the weaknesses of existing solutions, we propose a novel defense method, Dual Gradient Pruning (DGP), based on gradient pruning, which can improve communication efficiency while preserving the utility and privacy of CL. Specifically, DGP slightly changes gradient pruning with a stronger privacy guarantee. And DGP can also significantly improve communication efficiency with a theoretical analysis of its convergence and generalization. Our extensive experiments show that DGP can effectively defend against the most powerful GIAs and reduce the communication cost without sacrificing the model's utility.

Document-level relation extraction (DocRE) poses the challenge of identifying relationships between entities within a document. Existing approaches rely on logical reasoning or contextual cues from entities. This paper reframes document-level RE as link prediction over a Knowledge Graph (KG) with distinct benefits: 1) Our approach amalgamates entity context and document-derived logical reasoning, enhancing link prediction quality. 2) Predicted links between entities offer interpretability, elucidating employed reasoning. We evaluate our approach on benchmark datasets - DocRED, ReDocRED, and DWIE. The results indicate that our proposed method outperforms the state-of-the-art models and suggests that incorporating context-based Knowledge Graph link prediction techniques can enhance the performance of document-level relation extraction models.

In representation learning, a disentangled representation is highly desirable as it encodes generative factors of data in a separable and compact pattern. Researchers have advocated leveraging disentangled representations to complete downstream tasks with encouraging empirical evidence. This paper further investigates the necessity of disentangled representation in downstream applications. Specifically, we show that dimension-wise disentangled representations are unnecessary on a fundamental downstream task, abstract visual reasoning. We provide extensive empirical evidence against the necessity of disentanglement, covering multiple datasets, representation learning methods, and downstream network architectures. Furthermore, our findings suggest that the informativeness of representations is a better indicator of downstream performance than disentanglement. Finally, the positive correlation between informativeness and disentanglement explains the claimed usefulness of disentangled representations in previous works. The source code is available at https://github.com/Richard-coder-Nai/disentanglement-lib-necessity.git

For addressing the data privacy and portability issues of domain adaptation, Domain Adaptation of Black-box Predictors (DABP) aims to adapt a black-box source model to an unlabeled target domain without accessing both the source-domain data and details of the source model. Although existing DABP approaches based on knowledge distillation (KD) have achieved promising results, we experimentally find that these methods all have the minority class forgetting issue, which refers that the trained model completely forgets some minority classes. To address this issue, we propose a method called Reviewing the Forgotten Classes (RFC), which including two main modules. Firstly, we propose a simple but effective component called selection training (ST). ST selects classes that the model tends to forget according to the learning status of the model and obtains clean samples of the selected classes with the small-loss criterion for enhanced training. ST is orthogonal to previous methods and can effectively alleviate their minority class forgetting issue. Secondly, we find that neighborhood clustering (NC) can help the model learn more balanced than KD so that further alleviate the minority class forgetting issue. However, NC is based on the fact that target features from the source model already form some semantic structure, while DABP is unable to obtain the source model. Thus, we use KD and ST to warm up the target model to form a certain semantic structure. Overall, our method inherits the merits of both ST and NC, and achieves state of the art on three DABP benchmarks.

Designed to establish potential relations and distill high-order representations, graph-based recommendation systems continue to reveal promising results by jointly modeling ratings and reviews. However, existing studies capture simple review relations, failing to (1) completely explore hidden connections between users (or items), (2) filter out redundant information derived from reviews, and (3) model the behavioral association between rating and review interactions. To address these challenges, we propose a review-enhanced hierarchical contrastive learning, namely ReHCL. First, ReHCL constructs topic and semantic graphs to fully mine review relations from different views. Moreover, a cross-view graph contrastive learning is used to achieve enhancement of node representations and extract useful review knowledge. Meanwhile, we design a neighbor-based positive sampling to capture the graph-structured similarity between topic and semantic views, further performing efficient contrast and reducing redundant noise. Next, we propose a cross-modal contrastive learning to match the rating and review representations, by exploring the association between ratings and reviews. Lastly, these two contrastive learning modes form a hierarchical contrastive learning task, which is applied to enhance the final recommendation task. Extensive experiments verify the superiority of ReHCL compared with state-of-the-arts.

With the proliferation of dialogic data across the Internet, the Dialogue Commonsense Multi-choice Question Answering (DC-MCQ) task has emerged as a response to the challenge of comprehending user queries and intentions. Although prevailing methodologies exhibit effectiveness in addressing single-choice questions, they encounter difficulties in handling multi-choice queries due to the heightened intricacy and informational density. In this paper, inspired by the human cognitive process of progressively excluding options, we propose a three-step Reverse Exclusion Graph-of-Thought (ReX-GoT) framework, including Option Exclusion, Error Analysis, and Combine Information. Specifically, our ReX-GoT mimics human reasoning by gradually excluding irrelevant options and learning the reasons for option errors to choose the optimal path of the GoT and ultimately infer the correct answer. By progressively integrating intricate clues, our method effectively reduces the difficulty of multi-choice reasoning and provides a novel solution for DC-MCQ. Extensive experiments on the CICERO and CICERO_v2 datasets validate the significant improvement of our approach on DC-MCQ task. On zero-shot setting, our model outperform the best baseline by 17.67% in terms of F1 score for the multi-choice task. Most strikingly, our GPT3.5-based ReX-GoT framework achieves a remarkable 39.44% increase in F1 score.

Compositional Zero-Shot Learning (CZSL) aims to transfer knowledge from seen state-object pairs to novel unseen pairs. In this process, visual bias caused by the diverse interrelationship of state-object combinations blurs their visual features, hindering the learning of distinguishable class prototypes. Prevailing methods concentrate on disentangling states and objects directly from visual features, disregarding potential enhancements that could arise from a data viewpoint. Experimentally, we unveil the results caused by the above problem closely approximate the long-tailed distribution. As a solution, we transform CZSL into a proximate class imbalance problem. We mathematically deduce the role of class prior within the long-tailed distribution in CZSL. Building upon this insight, we incorporate visual bias caused by compositions into the classifier's training and inference by estimating it as a proximate class prior. This enhancement encourages the classifier to acquire more discernible class prototypes for each composition, thereby achieving more balanced predictions. Experimental results demonstrate that our approach elevates the model's performance to the state-of-the-art level, without introducing additional parameters.

Machine learning-assisted retrosynthesis prediction models have been gaining widespread adoption, though their performances oftentimes degrade significantly when deployed in real-world applications embracing out-of-distribution (OOD) molecules or reactions. Despite steady progress on standard benchmarks, our understanding of existing retrosynthesis prediction models under the premise of distribution shifts remains stagnant. To this end, we first formally sort out two types of distribution shifts in retrosynthesis prediction and construct two groups of benchmark datasets. Next, through comprehensive experiments, we systematically compare state-of-the-art retrosynthesis prediction models on the two groups of benchmarks, revealing the limitations of previous in-distribution evaluation and re-examining the advantages of each model. More remarkably, we are motivated by the above empirical insights to propose two model-agnostic techniques that can improve the OOD generalization of arbitrary off-the-shelf retrosynthesis prediction algorithms. Our preliminary experiments show their high potential with an average performance improvement of 4.6%, and the established benchmarks serve as a foothold for further retrosynthesis prediction research towards OOD generalization.

Compositional zero-shot learning (CZSL) aims to recognize unseen attribute-object compositions by learning from seen compositions. Composing the learned knowledge of seen primitives, i.e., attributes or objects, into novel compositions is critical for CZSL. In this work, we propose to explicitly retrieve knowledge of seen primitives for compositional zero-shot learning. We present a retrieval-augmented method, which augments standard multi-path classification methods with two retrieval modules. Specifically, we construct two databases storing the attribute and object representations of training images, respectively. For an input training/testing image, we use two retrieval modules to retrieve representations of training images with the same attribute and object, respectively. The primitive representations of the input image are augmented by using the retrieved representations, for composition recognition. By referencing semantically similar images, the proposed method is capable of recalling knowledge of seen primitives for compositional generalization. Experiments on three widely-used datasets show the effectiveness of the proposed method.

Face retouching is to beautify a face image, while preserving the image content as much as possible. It is a promising yet challenging task to remove face imperfections and fill with normal skin. Generic image enhancement methods are hampered by the lack of imperfection localization, which often results in incomplete removal of blemishes at large scales. To address this issue, we propose a transformer-based approach, RetouchFormer, which simultaneously identify imperfections and synthesize realistic content in the corresponding regions. Specifically, we learn a latent dictionary to capture the clean face priors, and predict the imperfection regions via a reconstruction-oriented localization module. Also based on this, we can realize face retouching by explicitly suppressing imperfections in our selective self-attention computation, such that local content will be synthesized from normal skin. On the other hand, multi-scale feature tokens lead to increased flexibility in dealing with the imperfections at various scales. The design elements bring greater effectiveness and efficiency. RetouchFormer outperforms the advanced face retouching methods and synthesizes clean face images with high fidelity in our list of extensive experiments performed.