IJCAI 2023 Papers

Sharing intentions is crucial for efficient cooperation in communication-enabled multi-agent reinforcement learning. Recent work applies static or undirected graphs to determine the order of interaction. However, the static graph is not general for complex cooperative tasks, and the parallel message-passing update in the undirected graph with cycles cannot guarantee convergence. To solve this problem, we propose Deep Hierarchical Communication Graph (DHCG) to learn the dependency relationships between agents based on their messages. The relationships are formulated as directed acyclic graphs (DAGs), where the selection of the proper topology is viewed as an action and trained in an end-to-end fashion. To eliminate the cycles in the graph, we apply an acyclicity constraint as intrinsic rewards and then project the graph in the admissible solution set of DAGs. As a result, DHCG removes redundant communication edges for cost improvement and guarantees convergence. To show the effectiveness of the learned graphs, we propose policy-based and value-based DHCG. Policy-based DHCG factorizes the joint policy in an auto-regressive manner, and value-based DHCG factorizes the joint value function to individual value functions and pairwise payoff functions. Empirical results show that our method improves performance across various cooperative multi-agent tasks, including Predator-Prey, Multi-Agent Coordination Challenge, and StarCraft Multi-Agent Challenge.

In financial scenarios, influenced by common factors such as global macroeconomic and sector-specific factors, stocks exhibit varying degrees of correlations with each other, which is essential in risk-averse portfolio allocation. Because the real risk matrix is unobservable, the covariance-based correlation matrix is widely used for constructing diversified stock portfolios. However, studies have seldom focused on dynamic correlation matrix estimation under the non-stationary financial market. Moreover, as the number of stocks in the market grows, existing correlation matrix estimation methods face more serious challenges with regard to efficiency and effectiveness. In this paper, we propose a novel hash-based dynamic correlation forecasting model (HDCF) to estimate dynamic stock correlations. Under structural assumptions on the correlation matrix, HDCF learns the hash representation based on normalizing flows instead of the real-valued representation, which performs extremely efficiently in high-dimensional settings. Experiments show that our proposed model outperforms baselines on portfolio decisions in terms of effectiveness and efficiency.

Semi-supervised semantic segmentation methods are the main solution to alleviate the problem of high annotation consumption in semantic segmentation. However, the class imbalance problem makes the model favor the head classes with sufficient training samples, resulting in poor performance of the tail classes. To address this issue, we propose a Decoupled Semi-Supervise Semantic Segmentation (DeS4) framework based on the teacher-student model. Specifically, we first propose a decoupling training strategy to split the training of the encoder and segmentation decoder, aiming at a balanced decoder. Then, a non-learnable prototype-based segmentation head is proposed to regularize the category representation distribution consistency and perform a better connection between the teacher model and the student model. Furthermore, a Multi-Entropy Sampling (MES) strategy is proposed to collect pixel representation for updating the shared prototype to get a class-unbiased head. We conduct extensive experiments of the proposed DeS4 on two challenging benchmarks (PASCAL VOC 2012 and Cityscapes) and achieve remarkable improvements over the previous state-of-the-art methods.

Recent studies have proposed models that yielded promising performance for the hateful meme classification task. Nevertheless, these proposed models do not generate interpretable explanations that uncover the underlying meaning and support the classification output. A major reason for the lack of explainable hateful meme methods is the absence of a hateful meme dataset that contains ground truth explanations for benchmarking or training. Intuitively, having such explanations can educate and assist content moderators in interpreting and removing flagged hateful memes. This paper address this research gap by introducing Hateful meme with Reasons Dataset (HatReD), which is a new multimodal hateful meme dataset annotated with the underlying hateful contextual reasons. We also define a new conditional generation task that aims to automatically generate underlying reasons to explain hateful memes and establish the baseline performance of state-of-the-art pre-trained language models on this task. We further demonstrate the usefulness of HatReD by analyzing the challenges of the new conditional generation task in explaining memes in seen and unseen domains. The dataset and benchmark models are made available here: https://github.com/Social-AI-Studio/HatRed

We consider the problem of detecting adversarial attacks against cooperative multi-agent reinforcement learning. We propose a decentralized scheme that allows agents to detect the abnormal behavior of one compromised agent. Our approach is based on a recurrent neural network (RNN) trained during cooperative learning to predict the action distribution of other agents based on local observations. The predicted distribution is used for computing a normality score for the agents, which allows the detection of the misbehavior of other agents. To explore the robustness of the proposed detection scheme, we formulate the worst-case attack against our scheme as a constrained reinforcement learning problem. We propose to compute an attack policy by optimizing the corresponding dual function using reinforcement learning. Extensive simulations on various multi-agent benchmarks show the effectiveness of the proposed detection scheme in detecting state-of-the-art attacks and in limiting the impact of undetectable attacks.

Time series forecasting is prevalent in various real-world applications. Despite the promising results of deep learning models in time series forecasting, especially the Recurrent Neural Networks (RNNs), the explanations of time series models, which are critical in high-stakes applications, have received little attention. In this paper, we propose a Decomposition-based Linear Explainable LSTM (DeLELSTM) to improve the interpretability of LSTM. Conventionally, the interpretability of RNNs only concentrates on the variable importance and time importance. We additionally distinguish between the instantaneous influence of new coming data and the long-term effects of historical data. Specifically, DeLELSTM consists of two components, i.e., standard LSTM and tensorized LSTM. The tensorized LSTM assigns each variable with a unique hidden state making up a matrix h(t), and the standard LSTM models all the variables with a shared hidden state H(t). By decomposing the H(t) into the linear combination of past information h(t-1) and the fresh information h(t)-h(t-1), we can get the instantaneous influence and the long-term effect of each feature. In addition, the advantage of linear regression also makes the explanation transparent and clear. We demonstrate the effectiveness and interpretability of DeLELSTM on three empirical datasets. Extensive experiments show that the proposed method achieves competitive performance against the baseline methods and provides a reliable explanation relative to domain knowledge.

The article proposes a new approach to reasoning about knowledge and strategies in multiagent systems. It emphasizes data, not agents, as the source of strategic knowledge. The approach brings together Armstrong's functional dependency expression from database theory, a data-informed knowledge modality based on a recent work by Baltag and van Benthem, and a newly proposed data-informed strategy modality. The main technical result is a sound and complete logical system that describes the interplay between these three logical operators.

In multi-agent systems, norm enforcement is a mechanism for steering the behavior of individual agents in order to achieve desired system-level objectives. Due to the dynamics of multi-agent systems, however, it is hard to design norms that guarantee the achievement of the objectives in every operating context. Also, these objectives may change over time, thereby making previously defined norms ineffective. In this paper, we investigate the use of system execution data to automatically synthesise and revise conditional prohibitions with deadlines, a type of norms aimed at preventing agents from exhibiting certain patterns of behaviors. We propose DDNR (Data-Driven Norm Revision), a data-driven approach to norm revision that synthesises revised norms with respect to a data set of traces describing the behavior of the agents in the system. We evaluate DDNR using a state-of-the-art, off-the-shelf urban traffic simulator. The results show that DDNR synthesises revised norms that are significantly more accurate than the original norms in distinguishing adequate and inadequate behaviors for the achievement of the system-level objectives.

The weakest pre-expectation framework from Morgan and McIver for deductive verification of probabilistic programs generalizes binary state assertions to real-valued expectations to measure expected values of expressions over probabilistic program variables. While loop-free programs can be analyzed by mechanically transforming expectations, verifying programs with loops requires finding an invariant expectation. We view invariant expectation synthesis as a regression problem: given an input state, predict the average value of the post-expectation in the output distribution. With this perspective, we develop the first data-driven invariant synthesis method for probabilistic programs. Unlike prior work on probabilistic invariant inference, our approach learns piecewise continuous invariants without relying on template expectations. We also develop a data-driven approach to learn sub-invariants from data, which can be used to upper- or lower-bound expected values. We implement our approaches and demonstrate their effectiveness on a variety of benchmarks from the probabilistic programming literature.

The conventional lottery ticket hypothesis (LTH) claims that there exists a sparse subnetwork within a dense neural network and a proper random initialization method, called the winning ticket, such that it can be trained from scratch to almost as good as the dense counterpart. Meanwhile, the research of LTH in vision transformers (ViTs) is scarcely evaluated. In this paper, we first show that the conventional winning ticket is hard to find at weight level of ViTs by existing methods. Then, we generalize the LTH for ViTs to input data consisting of image patches inspired by the input dependence of ViTs. That is, there exists a subset of input image patches such that a ViT can be trained from scratch by using only this subset of patches and achieve similar accuracy to the ViTs trained by using all image patches. We call this subset of input patches the winning tickets, which represent a significant amount of information in the input data. We use a ticket selector to generate the winning tickets based on the informativeness of patches for various types of ViT, including DeiT, LV-ViT, and Swin Transformers. The experiments show that there is a clear difference between the performance of models trained with winning tickets and randomly selected subsets, which verifies our proposed theory. We elaborate the analogical similarity between our proposed Data-LTH-ViTs and the conventional LTH for further verifying the integrity of our theory. The Source codes are available at https://github.com/shawnricecake/vit-lottery-ticket-input.

Communication lays the foundation for cooperation in human society and in multi-agent reinforcement learning (MARL). Humans also desire to maintain their privacy when communicating with others, yet such privacy concern has not been considered in existing works in MARL. We propose the differentially private multi-agent communication (DPMAC) algorithm, which protects the sensitive information of individual agents by equipping each agent with a local message sender with rigorous (epsilon, delta)-differential privacy (DP) guarantee. In contrast to directly perturbing the messages with predefined DP noise as commonly done in privacy-preserving scenarios, we adopt a stochastic message sender for each agent respectively and incorporate the DP requirement into the sender, which automatically adjusts the learned message distribution to alleviate the instability caused by DP noise. Further, we prove the existence of a Nash equilibrium in cooperative MARL with privacy-preserving communication, which suggests that this problem is game-theoretically learnable. Extensive experiments demonstrate a clear advantage of DPMAC over baseline methods in privacy-preserving scenarios.

Recently, techniques utilizing frequency-based methods have gained significant attention, as they exhibit exceptional restoration capabilities for detail and structure in video super-resolution tasks. However, most of these frequency-based methods mainly have three major limitations: 1) insufficient exploration of object motion information, 2) inadequate enhancement for high-fidelity regions, and 3) loss of spatial information during convolution. In this paper, we propose a novel network, Directional Frequency Video Super-Resolution (DFVSR), to address these limitations. Specifically, we reconsider object motion from a new perspective and propose Directional Frequency Representation (DFR), which not only borrows the property of frequency representation of detail and structure information but also contains the direction information of the object motion that is extremely significant in videos. Based on this representation, we propose a Directional Frequency-Enhanced Alignment (DFEA) to use double enhancements of task-related information for ensuring the retention of high-fidelity frequency regions to generate the high-quality alignment feature. Furthermore, we design a novel Asymmetrical U-shaped network architecture to progressively fuse these alignment features and output the final output. This architecture enables the intercommunication of the same level of resolution in the encoder and decoder to achieve the supplement of spatial information. Powered by the above designs, our method achieves superior performance over state-of-the-art models on both quantitative and qualitative evaluations.

Exploration is a fundamental aspect of reinforcement learning (RL), and its effectiveness is a deciding factor in the performance of RL algorithms, especially when facing sparse extrinsic rewards. Recent studies have shown the effectiveness of encouraging exploration with intrinsic rewards estimated from novelties in observations. However, there is a gap between the novelty of an observation and an exploration, as both the stochasticity in the environment and the agent's behavior may affect the observation. To evaluate exploratory behaviors accurately, we propose DEIR, a novel method in which we theoretically derive an intrinsic reward with a conditional mutual information term that principally scales with the novelty contributed by agent explorations, and then implement the reward with a discriminative forward model. Extensive experiments on both standard and advanced exploration tasks in MiniGrid show that DEIR quickly learns a better policy than the baselines. Our evaluations on ProcGen demonstrate both the generalization capability and the general applicability of our intrinsic reward.

In the realm of autonomous driving, real-time perception or streaming perception remains under-explored. This research introduces DAMO-StreamNet, a novel framework that merges the cutting-edge elements of the YOLO series with a detailed examination of spatial and temporal perception techniques. DAMO-StreamNet's main inventions include: (1) a robust neck structure employing deformable convolution, bolstering receptive field and feature alignment capabilities; (2) a dual-branch structure synthesizing short-path semantic features and long-path temporal features, enhancing the accuracy of motion state prediction; (3) logits-level distillation facilitating efficient optimization, which aligns the logits of teacher and student networks in semantic space; and (4) a real-time prediction mechanism that updates the features of support frames with the current frame, providing smooth streaming perception during inference. Our testing shows that DAMO-StreamNet surpasses current state-of-the-art methodologies, achieving 37.8% (normal size (600, 960)) and 43.3% (large size (1200, 1920)) sAP without requiring additional data. This study not only establishes a new standard for real-time perception but also offers valuable insights for future research. The source code is at https://github.com/zhiqic/DAMO-StreamNet.

Accurate prediction of crop yield under the conditions of climate change is crucial to ensure food security. Transformers have shown remarkable success in modeling sequential data and hold the potential for improving crop yield prediction. To understand how weather and meteorological sequence variables affect crop yield, the positional encoding used in Transformers is typically shared across different sample sequences. We argue that it is necessary and beneficial to differentiate the positional encoding for distinct samples based on time-invariant properties of the sequences. Particularly, the sequence variables influencing crop yield vary according to static variables such as geographical locations. Sample data from southern areas may benefit from more tailored positional encoding different from that for northern areas. We propose a novel transformer based architecture for accurate and robust crop yield prediction, by introducing a Customized Positional Encoding (CPE) that encodes a sequence adaptively according to static information associated with the sequence. Empirical studies demonstrate the effectiveness of the proposed novel architecture and show that partially lin- earized attention better captures the bias introduced by side information than softmax re-weighting. The resultant crop yield prediction model is robust to climate change, with mean-absolute-error reduced by up to 26% compared to the best baseline model in extreme drought years.

In large-scale e-commerce live-stream recommendation, streamers are classified into different levels based on their popularity and other metrics for marketing. Several top streamers at the head level occupy a considerable amount of exposure, resulting in an unbalanced data distribution. A unified model for all levels without consideration of imbalance issue can be biased towards head streamers and neglect the conflicts between levels. The lack of inter-level streamer correlations and intra-level streamer characteristics modeling imposes obstacles to estimating the user behaviors. To tackle these challenges, we propose a curriculum multi-level learning framework for imbalanced recommendation. We separate model parameters into shared and level-specific ones to explore the generality among all levels and discrepancy for each level respectively. The level-aware gradient descent and a curriculum sampling scheduler are designed to capture the de-biased commonalities from all levels as the shared parameters. During the specific parameters training, the hardness-aware learning rate and an adaptor are proposed to dynamically balance the training process. Finally, shared and specific parameters are combined to be the final model weights and learned in a cooperative training framework. Extensive experiments on a live-stream production dataset demonstrate the superiority of the proposed framework.

Graph machine learning has been extensively studied in both academia and industry. However, in the literature, most existing graph machine learning models are designed to conduct training with data samples in a random order, which may suffer from suboptimal performance due to ignoring the importance of different graph data samples and their training orders for the model optimization status. To tackle this critical problem, curriculum graph machine learning (Graph CL), which integrates the strength of graph machine learning and curriculum learning, arises and attracts an increasing amount of attention from the research community. Therefore, in this paper, we comprehensively overview approaches on Graph CL and present a detailed survey of recent advances in this direction. Specifically, we first discuss the key challenges of Graph CL and provide its formal problem definition. Then, we categorize and summarize existing methods into three classes based on three kinds of graph machine learning tasks, i.e., node-level, link-level, and graph-level tasks. Finally, we share our thoughts on future research directions. To the best of our knowledge, this paper is the first survey for curriculum graph machine learning.

Cross-community learning incorporates data from different sources to leverage task-specific solutions in a target community. This approach is particularly interesting for low-resource or newly created online communities, where data formalizing interactions between agents (community members) are limited. In such scenarios, a normative system that intends to regulate online interactions faces the challenge of continuously learning the meaning of norm violation as communities' views evolve, either with changes in the understanding of what it means to violate a norm or with the emergence of new violation classes. To address this issue, we propose the Cross-community Adapter Learning (CAL) framework, which combines adapters and transformer-based models to learn the meaning of norm violations expressed as textual sentences. Additionally, we analyze the differences in the meaning of norm violations between communities, using Integrated Gradients (IG) to understand the inner workings of our model and calculate a global relevance score that indicates the relevance of words for violation detection. Results show that cross-community learning enhances CAL's performance while explaining the differences in the meaning of norm-violating behavior based on community members' feedback. We evaluate our proposal in a small set of interaction data from Wikipedia, in which the norm prohibits hate speech.

Audio-visual speech recognition (AVSR) research has gained a great success recently by improving the noise-robustness of audio-only automatic speech recognition (ASR) with noise-invariant visual information. However, most existing AVSR approaches simply fuse the audio and visual features by concatenation, without explicit interactions to capture the deep correlations between them, which results in sub-optimal multimodal representations for downstream speech recognition task. In this paper, we propose a cross-modal global interaction and local alignment (GILA) approach for AVSR, which captures the deep audio-visual (A-V) correlations from both global and local perspectives. Specifically, we design a global interaction model to capture the A-V complementary relationship on modality level, as well as a local alignment approach to model the A-V temporal consistency on frame level. Such a holistic view of cross-modal correlations enable better multimodal representations for AVSR. Experiments on public benchmarks LRS3 and LRS2 show that our GILA outperforms the supervised learning state-of-the-art. Code is at https://github.com/YUCHEN005/GILA.