AAAI 2025 论文列表

With the growing demand for solutions to real-world video challenges, interest in dense video captioning (DVC) has been on the rise. DVC involves the automatic captioning and localization of untrimmed videos. Several studies highlight the challenges of DVC and introduce improved methods utilizing prior knowledge such as pre-training and external memory. In this research, we propose a model that leverages the prior knowledge of human-oriented hierarchical dense memory inspired by human memory hierarchy and cognition. To mimic human-like memory recall, we construct a hierarchical memory and a hierarchical memory reading module. We build an efficient hierarchical dense memory by employing clustering of memory events and summarization using large language models. Comparative experiments demonstrate that this hierarchical memory recall process improves the performance of DVC by achieving state-of-the-art performance on YouCook2 and ViTT datasets.

Existing knowledge distillation (KD) studies for streaming automatic speech recognition (ASR) adopt a non-streaming model as the teacher and a streaming model as the student, respectively. Since the non-streaming teacher usually has less emission latency compared to the streaming student, the teacher's prediction is typically shifted by $\tau$ frames, where the parameter $\tau$ is selected heuristically. In this paper, we observe that this manual shifting is sub-optimal and propose a novel framework, namely Heuristic-free KD. Instead of leveraging knowledge from the non-streaming teacher model, we employ a self-distillation setup, distilling the knowledge within the streaming architecture itself. Since the teacher and student share the same streaming ASR backbone, the alignment mismatch issue can be effectively mitigated without requiring any time shifting by $\tau$. Additionally, we incorporate full-context textual information as an auxiliary multi-modal input for the proposed teacher. Although the streaming architecture lacks future context, the additional linguistic input enables it to generate more accurate knowledge for self-distillation. We empirically demonstrate that the proposed KD approach significantly improves the performance of the streaming ASR model, outperforming conventional methods that rely on the offline teacher and heuristic parameter.

In cross-modal retrieval, comprehensive image understanding is vital while the scene text in images can provide fine-grained information to understand visual semantics. Current methods fail to make full use of scene text. They suffer from the semantic ambiguity of independent scene text and overlook the heterogeneous concepts in image-caption pairs. In this paper, we propose a heterogeneous prompt-guided entity inferring and distilling (HOPID) network to explore the nature connection of scene text in images and captions and learn a property-centric scene text representation. Specifically, we propose to align scene text in images and captions via heterogeneous prompt, which consists of visual and text prompt. For text prompt, we introduce the discriminative entity inferring module to reason key scene text words from captions, while visual prompt highlights the corresponding scene text in images. Furthermore, to secure a robust scene text representation, we design a perceptive entity distilling module that distills the beneficial information of scene text at a fine-grained level. Extensive experiments show that the proposed method significantly outperforms existing approaches on two public cross-modal retrieval benchmarks.

Multi-Robot Coverage problems have been extensively studied in robotics, planning and multi-agent systems. In this work, we consider the coverage problem when there are constraints on the proximity (e.g., maximum distance between the agents, or a blue agent must be adjacent to a red agent) and the movement (e.g., terrain traversability and material load capacity) of the robots. Such constraints naturally arise in many real-world applications, e.g. in search-and-rescue and maintenance operations. Given such a setting, the goal is to compute a covering tour of the graph with a minimum number of steps, and that adheres to the proximity and movement constraints. For this problem, our contributions are four: (i) a formal formulation of the problem, (ii) an exact algorithm that is FPT in parameters ||F||, d and ω - the set of robot formations that encode the proximity constraints, the maximum nodes degree, and the tree-width of the graph, respectively, (iii) for the case that the graph is a tree: a PTAS approximation scheme, that given an ε produces a tour that is within a 1+ ε⋅error(||F||, d)) of the optimal one, and the computation runs in time poly(n) ⋅ h(1/ε, ||F||). (iv) for the case that the graph is a tree, with k=3 robots, and the constraint is that all agents are connected: a PTAS scheme with multiplicative approximation error of 1 + O(ε), independent of d.

We study a hinted heterogeneous multi-agent multi-armed bandits problem (HMA2B), where agents can query low-cost observations (hints) in addition to pulling arms. In this framework, each of the M agents has a unique reward distribution over K arms, and in T rounds, they can observe the reward of the arm they pull only if no other agent pulls that arm. The goal is to maximize the total utility by querying the minimal necessary hints without pulling arms, achieving time-independent regret. We study HMA2B in both centralized and decentralized setups. Our main centralized algorithm, GP-HCLA, which is an extension of HCLA, uses a central decision-maker for arm-pulling and hint queries, achieving O(M^4 K) regret with O(M K log T) adaptive hints. In decentralized setups, we propose two algorithms, HD-ETC and EBHD-ETC, that allow agents to choose actions independently through collision-based communication and query hints uniformly until stopping, yielding O(M^3 K^2) regret with O(M^3 K log T) hints, where the former requires knowledge of the minimum gap and the latter does not. Finally, we establish lower bounds to prove the optimality of our results and verify them through numerical simulations.

The recent past has seen an increasing interest in Heterogeneous Graph Neural Networks (HGNNs), since many real-world graphs are heterogeneous in nature, from citation graphs to email graphs. However, existing methods ignore a tree hierarchy among metapaths, naturally constituted by different node types and relation types. In this paper, we present HetTree, a novel HGNN that models both the graph structure and heterogeneous aspects in a scalable and effective manner. Specifically, HetTree builds a semantic tree data structure to capture the hierarchy among metapaths. To effectively encode the semantic tree, HetTree uses a novel subtree attention mechanism to emphasize metapaths that are more helpful in encoding parent-child relationships. Moreover, HetTree proposes carefully matching pre-computed features and labels correspondingly, constituting a complete metapath representation. Our evaluation of HetTree on a variety of real-world datasets demonstrates that it outperforms all existing baselines on open benchmarks and efficiently scales to large real-world graphs with millions of nodes and edges.

The challenges tied to unstructured graph data are manifold, primarily falling into node, edge, and graph-level problem categories. Graph Neural Networks (GNNs) serve as effective tools to tackle these issues. However, individual tasks often demand distinct model architectures, and training these models typically requires abundant labeled data, a luxury often unavailable in practical settings. Recently, various "prompt tuning" methodologies have emerged to empower GNNs to adapt to multi-task learning with limited labels. The crux of these methods lies in bridging the gap between pre-training tasks and downstream objectives. Nonetheless, a prevalent oversight in existing studies is the homophily-centric nature of prompt tuning frameworks, disregarding scenarios characterized by high heterogeneity. To remedy this oversight, we introduce a novel prompting strategy named HeterGP tailored for highly heterophilic scenarios. Specifically, we present a dual-view approach to capture both homophilic and heterophilic information, along with a prompt graph design that encompasses token initialization and insertion patterns. Through extensive experiments conducted in a few-shot context encompassing node and graph classification tasks, our method showcases superior performance in highly heterophilic environments compared to state-of-the-art prompt tuning techniques.

In forecasting competitions, the traditional mechanism scores the predictions of each contestant against the outcome of each event, and the contestant with the highest total score wins. While it is well-known that this traditional mechanism can suffer from incentive issues, it is folklore that contestants will still be roughly truthful as the number of events grows. Yet thus far the literature lacks a formal analysis of this traditional mechanism. This paper gives the first such analysis. We first demonstrate that the "long-run truthfulness" folklore is false: even for arbitrary numbers of events, the best forecaster can have an incentive to hedge, reporting more moderate beliefs to increase their win probability. On the positive side, however, we show that two contestants will be approximately truthful when they have sufficient uncertainty over the relative quality of their opponent and the outcomes of the events, a case which may arise in practice.

Heterogeneous graphs, which are common in real-world downstream tasks, have recently sparked a wave of research interest. The performance of end-to-end heterogeneous graph neural networks (HGNNs) greatly relies on supervised training for specific tasks. To reduce the labeling cost, the "pretrain-finetune" paradigm has been widely adopted, but it leads to a knowledge gap between the pre-trained model and downstream tasks. In an effort to address this gap, the "pretrain-prompt" paradigm has emerged as a promising approach. This involves fine-tuning randomly initialized learnable vectors in downstream tasks. However, this approach may result in an insufficient representation of downstream task features. Existing techniques for heterogeneous graph prompting restructure the heterogeneous graph to align with the homogeneous graph prompting scheme. This can potentially introduce the same limitations as homogeneous graph prompt learning. In this paper, we propose HePa, short for Heterogeneous Graph Prompting for all-level classification tasks. It not only includes a unified prompt template-graph adapted for heterogeneous graphs but also introduces a novel pre-prompt token optimized during the pre-training phase to convey task information downstream. With these designs, HePa can complete all levels of classification tasks toward few-shot scenarios while activating in-context learning. Finally, we conducted a comprehensive experimental analysis of HePa on three benchmark datasets.

Understanding and leveraging the 3D structures of proteins is central to a variety of biological and drug discovery tasks. While deep learning has been applied successfully for structure-based protein function prediction tasks, current methods usually employ distinct training for each task. However, each of the tasks is of small size, and such a single-task strategy hinders the models' performance and generalization ability. As some labeled 3D protein datasets are biologically related, combining multi-source datasets for larger-scale multi-task learning is one way to overcome this problem. In this paper, we propose a neural network model to address multiple tasks jointly upon the input of 3D protein structures. In particular, we first construct a standard structure-based multi-task benchmark called Protein-MT, consisting of 6 biologically relevant tasks, including affinity prediction and property prediction, integrated from 4 public datasets. Then, we develop a novel graph neural network for multi-task learning, dubbed Heterogeneous Multichannel Equivariant Network (HeMeNet), which is E(3) equivariant and able to capture heterogeneous relationships between different atoms. Besides, HeMeNet can achieve task-specific learning via the task-aware readout mechanism. Extensive evaluations of our benchmark verify the effectiveness of multi-task learning, and our model generally surpasses state-of-the-art models.

Table Understanding (TU) has achieved promising advancements, but it faces the challenges of the scarcity of manually labeled tables and the presence of complex table structures. To address these challenges, we propose HeGTa, a heterogeneous graph (HG)-enhanced large language model (LLM) designed for few-shot TU tasks. This framework aligns structural table semantics with the LLM's parametric knowledge through soft prompts and instruction tuning. It also addresses complex tables with a multi-task pre-training scheme, incorporating three novel multi-granularity self-supervised HG pre-text tasks. We empirically demonstrate the effectiveness of HeGTa, showing that it outperforms the SOTA for few-shot complex TU on several benchmarks.

The prevalence of multi-modal content on social media complicates automated moderation strategies. This calls for an enhancement in multi-modal classification and a deeper understanding of understated meanings in images and memes. Although previous efforts have aimed at improving model performance through fine-tuning, few have explored an end-to-end optimization pipeline that accounts for modalities, prompting, labelling, and fine-tuning. In this study, we propose an end-to-end conceptual framework for model opti- mization in complex tasks. Experiments support the efficacy of this traditional yet novel framework, achieving the highest accuracy and AUROC. Ablation experiments demonstrate that isolated optimisations are not ineffective on their own.

Machine learning techniques are notably vulnerable to natural or adversarial perturbations, which can lead to catastrophic failures with significant economic, ethical, and societal risks. In this New Faculty Highlight Talk, I will showcase my research on harnessing robust statistics to build robust and trustworthy AI systems. Specifically, I will highlight my research breakthroughs in graph learning (GNNs), large language models (LLMs), deep equilibrium models (DEQs), and general deep representation learning. These breakthroughs stem from a unified and principled robust statistics framework that incorporates robustness as the core inductive bias in deep learning architecture. This approach has enabled significant improvements in intrinsic robustness and generalization, even in complex and challenging environments. My research demonstrates the transformative potential of harnessing robust statistics in enhancing the robustness and trustworthiness of AI systems. Looking forward, I will continue to push this frontier by advocating the design of robustness-informed neural networks across various areas.

Recent advances in Large Language Models (LLMs) have demonstrated significant potential in the field of Recommendation Systems (RSs). Most existing studies have focused on converting user behavior logs into textual prompts and leveraging techniques such as prompt tuning to enable LLMs for recommendation tasks. Meanwhile, research interest has recently grown in multimodal recommendation systems that integrate data from images, text, and other sources using modality fusion techniques. This introduces new challenges to the existing LLM-based recommendation paradigm which relies solely on text modality information. Moreover, although Multimodal Large Language Models (MLLMs) capable of processing multi-modal inputs have emerged, how to equip MLLMs with multi-modal recommendation capabilities remains largely unexplored. To this end, in this paper, we propose the Multimodal Large Language Model-enhanced Sequential Multimodal Recommendation (MLLM-MSR) model. To capture the dynamic user preference, we design a two-stage user preference summarization method. Specifically, we first utilize an MLLM-based item-summarizer to extract image feature given an item and convert the image into text. Then, we employ a recurrent user preference summarization generation paradigm to capture the dynamic changes in user preferences based on an LLM-based user-summarizer. Finally, to enable the MLLM for multi-modal recommendation task, we propose to fine-tune a MLLM-based recommender using Supervised Fine-Tuning (SFT) techniques. Extensive evaluations across various datasets validate the effectiveness of MLLM-MSR, showcasing its superior ability to capture and adapt to the evolving dynamics of user preferences.

Large Language Models (LLMs) demonstrate remarkable capabilities, yet struggle with hallucination and outdated knowledge when tasked with complex knowledge reasoning, resulting in factually incorrect outputs. Previous studies have attempted to mitigate it by retrieving factual knowledge from large-scale knowledge graphs (KGs) to assist LLMs in logical reasoning and prediction of answers. However, this kind of approach often introduces noise and irrelevant data, especially in situations with extensive context from multiple knowledge aspects. In this way, LLM attention can be potentially mislead from question and relevant information. In our study, we introduce an Adaptive Multi-Aspect Retrieval-augmented over KGs (Amar) framework. This method retrieves knowledge including entities, relations, and subgraphs, and converts each piece of retrieved text into prompt embeddings. The Amar framework comprises two key sub-components: 1) a self-alignment module that aligns commonalities among entities, relations, and subgraphs to enhance retrieved text, thereby reducing noise interference; 2) a relevance gating module that employs a soft gate to learn the relevance score between question and multi-aspect retrieved data, to determine which information should be used to enhance LLMs' output, or even filtered altogether. Our method has achieved state-of-the-art performance on two common datasets, WebQSP and CWQ, showing a 1.9% improvement in accuracy over its best competitor and a 6.6% improvement in logical form generation over a method that directly uses retrieved text as context prompts. These results demonstrate the effectiveness of Amar in improving the reasoning of LLMs.

Heterogeneous graphs (HGs) that contain various node and edge types are ubiquitous in real-world scenarios. Considering the common label sparsity problem in HGs, some researchers propose to pretrain on source HGs to extract general knowledge and then fine-tune on a target HG for knowledge transfer. However, existing methods often assume that source and target HGs share a single heterogeneity, meaning that they have the same types of nodes and edges, which contradicts the real-world scenarios requiring cross-heterogeneity transfer. Although a recent study has made some preliminary attempts in cross-heterogeneity learning, its definition of general knowledge heavily rely on human knowledge, which lacks flexibility and further leads to a suboptimal transfer. To address the problem, we propose a novel Language Model-enhanced Cross-Heterogeneity learning model, namely LMCH. Specifically, we first design a metapath-based corpus construction method to unify HG representations as languages. The corpora of source HGs are then used to fine-tune a pretrained Language Model (LM), enabling the LM to autonomously extract general knowledge across different HGs. Furthermore, to fully utilize the extensive unlabeled nodes in a few-labeled target HG, we propose an iterative training pipeline with the help of an extra Graph Neural Network (GNN) predictor, enhanced by LM-GNN contrastive alignment at the end of each iteration. Extensive experiments on four real-world datasets have demonstrated the superior performance of LMCH over state-of-the-art methods.

In this paper, we draw an analogy between processing natural languages and processing multivariate event streams from vehicles in order to predict when and what error pattern is most likely to occur in the future for a given car. Our approach leverages the temporal dynamics and contextual relationships of our event data from a fleet of cars. Event data is composed of discrete values of error codes as well as continuous values such as time and mileage. Modelled by two causal Transformers, we can anticipate vehicle failures and malfunctions before they happen. Thus, we introduce CarFormer, a Transformer model trained via a new self-supervised learning strategy, and EPredictor, an autoregressive Transformer decoder model capable of predicting when and what error pattern will most likely occur after some error code apparition. Despite the challenges of high cardinality of event types, their unbalanced frequency of appearance and limited labelled data, our experimental results demonstrate the excellent predictive ability of our novel model. Specifically, with sequences of 160 error codes on average, our model is able with only half of the error codes to achieve 80% F1 score for predicting what error pattern will occur and achieves an average absolute error of 58.4 ± 13.2h when forecasting the time of occurrence, thus enabling confident predictive maintenance and enhancing vehicle safety.

In a surge of text-to-image (T2I) models and their customization methods that generate new images of a user-provided subject, current works focus on alleviating the costs incurred by a lengthy per-subject optimization. These zero-shot customization methods encode the image of a specified subject into a visual embedding which is then utilized alongside the textual embedding for diffusion guidance. The visual embedding incorporates intrinsic information about the subject, while the textual embedding provides a new context. However, the existing methods often 1) generate images with the same pose as an input image, and 2) exhibit deterioration in the subject's identity when facing a pose variation prompt. We first pin down the problem and show that redundant pose information in the visual embedding interferes with the pose indication in the textual embedding. Conversely, the textual embedding also harms the subject's identity which is tightly entangled with the pose in the visual embedding. As a remedy, we propose text-orthogonal visual embedding which effectively harmonizes with the given textual embedding. We also adopt the visual-only embedding and inject the subject's clear features utilizing a self-attention swap. Our method is both effective and robust, offering highly flexible zero-shot generation while effectively maintaining the subject's identity.

Creating group choreography from music is crucial in cultural entertainment and virtual reality, with a focus on generating harmonious movements. Despite growing interest, recent approaches often struggle with two major challenges: multi-dancer collisions and single-dancer foot sliding. To address these challenges, we propose a Trajectory-Controllable Diffusion (TCDiff) framework, which leverages non-overlapping trajectories to ensure coherent and aesthetically pleasing dance movements. To mitigate collisions, we introduce a Dance-Trajectory Navigator that generates collision-free trajectories for multiple dancers, utilizing a distance-consistency loss to maintain optimal spacing. Furthermore, to reduce foot sliding, we present a footwork adaptor that adjusts trajectory displacement between frames, supported by a relative forward-kinematic loss to further reinforce the correlation between movements and trajectories. Experiments demonstrate our method's superiority.