AAAI 2025 论文列表

Weakly supervised phrase grounding tasks aim to learn alignments between phrases and regions with coarse image-caption match information. One branch of previous methods established pseudo-label relationships between phrases and regions based on the Expectation-Maximization (EM) algorithm combined with contrastive learning. However, adopting a simplified batch-level local update (partial) of pseudo-labels in E-step is sub-optimal, while extending it to global update requires inefficiently numerous computations. In addition, their failure to consider potential false negative examples in contrastive loss negatively impacts the effectiveness of M-step optimization. To address these issues, we propose a Momentum Pseudo Labeling (MPL) method, which efficiently uses a momentum model to synchronize global pseudo-label updates on the fly with model parameter updating. Additionally, we explore potential relationships between phrases and regions from non-matching image-caption pairs and convert these false negative examples to positive ones in contrastive learning. Our approach achieved SOTA performance on 3 commonly used grounding datasets for weakly supervised phrase grounding tasks.

Text-to-Video generation, which utilizes the provided text prompt to generate high-quality videos, has drawn increasing attention and achieved great success due to the development of diffusion models recently. Existing methods mainly rely on a pre-trained text encoder to capture the semantic information and perform cross attention with the encoded text prompt to guide the generation of video. However, when it comes to complex prompts that contain dynamic scenes and multiple camera-view transformations, these methods can not decompose the overall information into separate scenes, as well as fail to smoothly change scenes based on the corresponding camera-views. To solve these problems, we propose a novel method, i.e., Modular-Cam. Specifically, to better understand a given complex prompt, we utilize a large language model to analyze user instructions and decouple them into multiple scenes together with transition actions. To generate a video containing dynamic scenes that match the given camera-views, we incorporate the widely-used temporal transformer into the diffusion model to ensure continuity within a single scene and propose CamOperator, a modular network based module that well controls the camera movements. Moreover, we propose AdaControlNet, which utilizes ControlNet to ensure consistency across scenes and adaptively adjusts the color tone of the generated video. Extensive qualitative and quantitative experiments prove our proposed Modular-Cam's strong capability of generating multi-scene videos together with its ability to achieve fine-grained control of camera movements. Generated results are available at https://modular-cam.github.io.

We study the problem of modeling a non-linear dynamical system when given a time series by deriving equations directly from the data. Despite the fact that time series data are given as input, models for dynamics and estimation algorithms that incorporate long-term temporal dependencies are largely absent from existing studies. In this paper, we introduce a latent state to allow time-dependent modeling and formulate this problem as a dynamics estimation problem in latent states. We face multiple technical challenges, including (1) modeling latent non-linear dynamics and (2) solving circular dependencies caused by the presence of latent states. To tackle these challenging problems, we propose a new method, Latent Non-Linear equation modeling (LaNoLem), that can model a latent non-linear dynamical system and a novel alternating minimization algorithm for effectively estimating latent states and model parameters. In addition, we introduce criteria to control model complexity without human intervention. Compared with the state-of-the-art model, LaNoLem achieves competitive performance for estimating dynamics while outperforming other methods in prediction.

Heterogeneity is a fundamental and challenging issue in federated learning, especially for the graph data due to the complex relationships among the graph nodes. To deal with the heterogeneity, lots of existing methods perform the weighted federation based on their calculated similarities between pairwise clients (i.e., subgraphs). However, their inter-subgraph similarities estimated with the outputs of local models are less reliable, because the final outputs of local models may not comprehensively represent the real distribution of subgraph data. In addition, they ignore the critical intra-heterogeneity which usually exists within each subgraph itself. To address these issues, we propose a novel Federated learning method by integrally modeling the Inter-Intra Heterogeneity (FedIIH). For the inter-subgraph relationship, we propose a novel hierarchical variational model to infer the whole distribution of subgraph data in a multi-level form, so that we can accurately characterize the inter-subgraph similarities with the global perspective. For the intra-heterogeneity, we disentangle the subgraph into multiple latent factors and partition the model parameters into multiple parts, where each part corresponds to a single latent factor. Our FedIIH not only properly computes the distribution similarities between subgraphs, but also learns disentangled representations that are robust to irrelevant factors within subgraphs, so that it successfully considers the inter- and intra- heterogeneity simultaneously. Extensive experiments on six homophilic and five heterophilic graph datasets in both non-overlapping and overlapping settings demonstrate the effectiveness of our method when compared with eight state-of-the-art methods. Specifically, FedIIH averagely outperforms the second-best method by a large margin of 5.79% on all heterophilic datasets.

Bayesian Optimization (BO) is a sample-efficient black-box optimizer commonly used in search spaces where hyperparameters are independent. However, in many practical AutoML scenarios, there will be dependencies among hyperparameters, forming a conditional search space, which can be partitioned into structurally distinct subspaces. The structure and dimensionality of hyperparameter configurations vary across these subspaces, challenging the application of BO. Some previous BO works have proposed solutions to develop multiple Gaussian Process models in these subspaces. However, these approaches tend to be inefficient as they require a substantial number of observations to guarantee each GP's performance and cannot capture relationships between hyperparameters across different subspaces. To address these issues, this paper proposes a novel approach to model the response surfaces of all subspaces in one, which can model the relationships between hyperparameters elegantly via a self-attention mechanism. Concretely, we design a structure-aware hyperparameter embedding to preserve the structural information. Then, we introduce an attention-based deep feature extractor, capable of projecting configurations with different structures from various subspaces into a unified feature space, where the response surfaces can be formulated using a single standard Gaussian Process. The empirical results on a simulation function, various real-world tasks, and HPO-B benchmark demonstrate that our proposed approach improves the efficacy and efficiency of BO within conditional search spaces.

Despite significant recent advances in the field of Deep Reinforcement Learning (DRL), such methods typically incur high cost of training to learn effective policies, thus posing cost and safety challenges in many practical applications. To improve the learning efficiency of (D)RL methods, transfer learning (TL) has emerged as a promising approach to leverage prior experience on a source domain to speed learning on a new, but related, target domain. In this paper, we take a novel model-informed approach to TL in DRL by assuming that we have knowledge of both the source and target domain models (which would be the case in the prevalent setting of DRL with simulators). While directly solving either the source or target MDP via solution methods like value iteration is computationally prohibitive, we exploit the fact that if the target and source MDPs differ only due to a small structural change in their rewards, we can apply structured value iteration methods in a procedure we term ModelDiff to solve the much smaller target-source ``Diff'' MDP for a reasonable horizon. This ModelDiff approach can then be integrated into extensions of standard DRL algorithms like ModelDiff (MD) DQN, where it provides enhanced provable lower bound guidance to DQN that often speeds convergence for the positive transfer case while critically avoiding decelerated learning in the negative transfer case. Experiments show that MD-DQN matches or outperforms existing TL methods and baselines in both positive and negative transfer settings.

As machine learning model modification techniques are extensively employed to obtain well-performing models at reduced costs, several studies have emerged to determine the presence of a modification relationship (i.e., lineage) between models. However, these methods are not robust to high-impact modification techniques and none of them have addressed the measurement of lineage closeness, which quantifies the degrees of modification. In this work, we visualize the changes in model decision boundaries resulting from different modification techniques and conclude that differences in decision boundaries serve as a precise metric of lineage closeness. Building upon this insight, we propose a modification-type agnostic and task-agnostic method to measure model lineage closeness by calculating mean adversarial distances from data points to decision boundaries and matching rate of data points, with data points selected through an efficient sampling method to reduce computational overhead. Moreover, we propose a novel indirect measurement approach to support lineage closeness measurement for models with different tasks. Finally, comprehensive experiments show that our design achieves an impressive 97% accuracy in lineage determination, and can precisely measure model lineage closeness for different modifications.

The Model AI Assignments session seeks to gather and disseminate the best assignment designs of the Artificial Intelligence (AI) Education community. Recognizing that assignments form the core of student learning experience, we here present abstracts of thirteen AI assignments from the 2025 session that are easily adoptable, playfully engaging, and flexible for a variety of instructor needs. Assignment specifications and supporting resources may be found at http://modelai.gettysburg.edu

Multimodal recommendation systems can learn users' preferences from existing user-item interactions as well as the semantics of multimodal data associated with items. Many existing methods model this through a multimodal user-item graph, approaching multimodal recommendation as a graph learning task. Graph Neural Networks (GNNs) have shown promising performance in this domain. Prior research has capitalized on GNNs' capability to capture neighborhood information within certain receptive fields (typically denoted by the number of hops, K) to enrich user and item semantics. We observe that the optimal receptive fields for GNNs can vary across different modalities. In this paper, we propose GNNs with Modality-Independent Receptive Fields, which employ separate GNNs with independent receptive fields for different modalities to enhance performance. Our results indicate that the optimal K for certain modalities on specific datasets can be as low as 1 or 2, which may restrict the GNNs' capacity to capture global information. To address this, we introduce a Sampling-based Global Transformer, which utilizes uniform global sampling to effectively integrate global information for GNNs. We conduct comprehensive experiments that demonstrate the superiority of our approach over existing methods.

Video scene detection involves assessing whether each shot and its surroundings belong to the same scene. Achieving this requires meticulously correlating multi-modal cues, e.g., visual entity and place modalities, among shots and comparing semantic changes around each shot. However, most methods treat multi-modal semantics equally and do not examine contextual differences between the two sides of a shot, leading to sub-optimal detection performance. In this paper, we propose the Modality-Aware Shot Relating and Comparing approach (MASRC), which enables relating shots per their own characteristics of visual entity and place modalities, as well as comparing multi-shots similarities to have scene changes explicitly encoded. Specifically, to fully harness the potential of visual entity and place modalities in modeling shot relations, we mine long-term shot correlations from entity semantics while simultaneously revealing short-term shot correlations from place semantics. In this way, we can learn distinctive shot features that consolidate coherence within scenes and amplify distinguishability across scenes. Once equipped with distinctive shot features, we further encode the relations between preceding and succeeding shots of each target shot by similarity convolution, aiding in the identification of scene ending shots. We validate the broad applicability of the proposed components in MASRC. Extensive experimental results on public benchmark datasets demonstrate that the proposed MASRC significantly advances video scene detection.

Mobility data from smartphones, connected cars, and GPS devices are widely used for tasks such as transportation mode classification and suspicious movement detection. Time series research, a closely related field, focuses more on classification methods. Yet, Mobility Data analysis faces unique challenges like geographic transferability and limited public data due to privacy issues. My PhD work focuses on developing reusable, interpretable MD representations. I created Trajectory Interval Forest and later Geolet, a shapelet-based transformation to improve MD classification across geographic regions. Ongoing research explores improving geographic transferability and event-based trajectory clustering.

Weakly Supervised Semantic Segmentation (WSSS) with image-level labels typically uses Class Activation Maps (CAM) to achieve dense predictions. Recently, Vision Transformer (ViT) has provided an alternative to generate localization maps from class-patch attention. However, due to insufficient constraints on modeling such attention, we observe that the Localization Attention Maps (LAM) often struggle with the artifact issue, i.e., patch regions with minimal semantic relevance are falsely activated by class tokens. In this work, we propose MoRe to address this issue and further explore the potential of LAM. Our findings suggest that imposing additional regularization on class-patch attention is necessary. To this end, we first view the attention as a novel directed graph and propose the Graph Category Representation module to implicitly regularize the interaction among class-patch entities. It ensures that class tokens dynamically condense the related patch information and suppress unrelated artifacts at a graph level. Second, motivated by the observation that CAM from classification weights maintains smooth localization of objects, we devise the Localization-informed Regularization module to explicitly regularize the class-patch attention. It directly mines the token relations from CAM and further supervises the consistency between class and patch tokens in a learnable manner. Extensive experiments on PASCAL VOC and MS COCO validate that MoRe effectively addresses the artifact issue and achieves state-of-the-art performance, surpassing recent single-stage and even multi-stage methods.

Recent advancements in Large Vision-Language Models (LVLMs) highlight their ability to integrate and process multi-modal information. However, hallucinations—where generated content is inconsistent with input vision and instructions—remain a challenge. In this paper, we analyze LVLMs' layer-wise decoding and identify that hallucinations can arise during the reasoning and factual information injection process. Additionally, as the number of generated tokens increases, the forgetting of the original prompt may also lead to hallucinations.To address this, we propose a training-free decoding method called Mixture of Layer Experts (MoLE). MoLE leverages a heuristic gating mechanism to dynamically select multiple layers of LVLMs as expert layers: the Final Expert, the Second Opinion expert, and the Prompt Retention Expert. By the cooperation of each expert, MoLE enhances the robustness and faithfulness of the generation process. Our extensive experiments demonstrate that MoLE significantly reduces hallucinations, outperforming the current state-of-the-art decoding techniques across three mainstream LVLMs and two established hallucination benchmarks. Moreover, our method reveals the potential of LVLMs to independently produce more reliable and accurate outputs.

Large Language Models (LLMs) are often English-centric due to the disproportionate distribution of languages in their pre-training data. Enhancing non-English language capabilities through post-pretraining often results in catastrophic forgetting of high-resource languages. Previous methods either achieve good expansion with severe forgetting or slight forgetting with poor expansion, indicating the challenge of balancing language expansion while preventing forgetting. In this paper, we propose a method called MoE-LPR (Mixture-of-Experts with Language Priors Routing) to alleviate this problem. MoE-LPR employs a two-stage training approach to enhance the multilingual capability. First, the model is post-pretrained into a Mixture-of-Experts(MoE) architecture by upcycling, where all the original parameters are frozen and new experts are added. In this stage, we focus improving the ability on expanded languages, without using any original language data. Then, the model reviews the knowledge of the original languages with replay data amounting to less than 1% of post-pretraining, where we incorporate language priors routing to better recover the abilities of the original languages. Evaluations on multiple benchmarks show that MoE-LPR outperforms other post-pretraining methods. Freezing original parameters preserves original language knowledge while adding new experts preserves the learning ability. Reviewing with LPR enables effective utilization of multilingual knowledge within the parameters. Additionally, the MoE architecture maintains the same inference overhead while increasing total model parameters. Extensive experiments demonstrate MoE-LPR’s effectiveness in improving expanded languages and preserving original language proficiency with superior scalability.

Conventional GAN-based models for talking head generation often suffer from limited quality and unstable training. Recent approaches based on diffusion models have attempted to address these limitations and improve fidelity. However, they still face challenges, such as intensive sampling times and difficulties in maintaining temporal consistency due to the high stochasticity of diffusion models. To overcome these challenges, we propose a novel motion-disentangled diffusion model for high-quality talking head generation, called MoDiTalker. We introduce two modules: the Audio-To-Motion (AToM) module, designed to generate synchronized lip movements from audio, and the Motion-To-Video (MToV) module, designed to produce high-quality talking head videos based on the generated motions. AToM excels in capturing subtle lip movements by leveraging an audio attention mechanism. Additionally, MToV enhances temporal consistency by utilizing an efficient tri-plane representation. Our experiments on standard benchmarks demonstrate that our model outperforms existing GAN-based and diffusion-based models. We also provide comprehensive ablation studies and user study results.

Perturbation-based mechanisms, such as differential privacy, mitigate gradient leakage attacks by introducing noise into the gradients, thereby preventing attackers from reconstructing clients' private data from the leaked gradients. However, can gradient perturbation protection mechanisms truly defend against all gradient leakage attacks? In this paper, we present the first attempt to break the shield of gradient perturbation protection in Federated Learning for the extraction of private information. We focus on common noise distributions, specifically Gaussian and Laplace, and apply our approach to DNN and CNN models. We introduce Mjölnir, a perturbation-resilient gradient leakage attack that is capable of removing perturbations from gradients without requiring additional access to the original model structure or external data. Specifically, we leverage the inherent diffusion properties of gradient perturbation protection to develop a novel diffusion-based gradient denoising model for Mjölnir. By constructing a surrogate client model that captures the structure of perturbed gradients, we obtain crucial gradient data for training the diffusion model. We further utilize the insight that monitoring disturbance levels during the reverse diffusion process can enhance gradient denoising capabilities, allowing Mjölnir to generate gradients that closely approximate the original, unperturbed versions through adaptive sampling steps. Extensive experiments demonstrate that Mjölnir effectively recovers the protected gradients and exposes the Federated Learning process to the threat of gradient leakage, achieving superior performance in gradient denoising and private data recovery.

Unified detection of digital and physical attacks in facial recognition systems has become a focal point of research in recent years. However, current multi-modal methods typically ignore the intra-class and inter-class variability across different types of attacks, leading to degraded performance. To address this limitation, we propose MoAE-CR, a framework that effectively leverages class-aware information for improved attack detection. Our improvements manifest at two levels, i.e., the feature and loss level. At the feature level, we propose Mixture-of-Attack-Experts (MoAEs) to capture more subtle differences among various types of fake faces. At the loss level, we introduce Class Regularization (CR) through the Disentanglement Module (DM) and the Cluster Distillation Module (CDM). The DM enhances class separability by increasing the distance between the centers of live and fake face classes. However, center-to-center constraints alone are insufficient to ensure distinctive representations for individual features. Thus, we propose the CDM to further cluster features around their class centers while maintaining separation from other classes. Moreover, specific attacks that significantly deviate from common attack patterns are often overlooked. To address this issue, our distance calculation prioritizes more distant features. Extensive experiments on two unified physical-digital attack datasets demonstrate the state-of-the-art performance of the proposed method.

We consider a general and realistic scenario involving non-stationary time series, consisting of several offline intervals with different distributions within a fixed offline time horizon, and an online interval that continuously receives new samples. For non-stationary time series, the data distribution in the current online interval may have appeared in previous offline intervals. We theoretically explore the feasibility of applying knowledge from offline intervals to the current online interval. To this end, we propose the Mixture of Online and Offline Experts (MOOE). MOOE learns static offline experts from offline intervals and maintains a dynamic online expert for the current online interval. It then adaptively combines the offline and online experts using a meta expert to make predictions for the samples received in the online interval. Specifically, we focus on theoretical analysis, deriving parameter convergence, regret bounds, and generalization error bounds to prove the effectiveness of the algorithm.

Literature reviews play a crucial role in scientific research for understanding the current state of research, identifying gaps, and guiding future studies on specific topics. However, the process of conducting a comprehensive literature review is yet time-consuming. This paper proposes a novel framework, collaborative knowledge minigraph agents (CKMAs), to automate scholarly literature reviews. A novel prompt-based algorithm, the knowledge minigraph construction agent (KMCA), is designed to identify relations between concepts from academic literature and automatically constructs knowledge minigraphs. By leveraging the capabilities of large language models on constructed knowledge minigraphs, the multiple path summarization agent (MPSA) efficiently organizes concepts and relations from different viewpoints to generate literature review paragraphs. We evaluate CKMAs on three benchmark datasets. Experimental results show the effectiveness of the proposed method, further revealing promising applications of LLMs in scientific research.