AAAI 2025 论文列表

Multimodal learning with incomplete modality is practical and challenging. Recently, researchers have focused on enhancing the robustness of pre-trained MultiModal Transformers (MMTs) under missing modality conditions by applying learnable prompts. However, these prompt-based methods face several limitations: (1) incomplete modalities provide restricted modal cues for task-specific inference, (2) dummy imputation for missing content causes information loss and introduces noise, and (3) static prompts are instance-agnostic, offering limited knowledge for instances with various missing conditions. To address these issues, we propose RAGPT, a novel Retrieval-AuGmented dynamic Prompt Tuning framework. RAGPT comprises three modules: (I) the multi-channel retriever, which identifies similar instances through a within-modality retrieval strategy, (II) the missing modality generator, which recovers missing information using retrieved contexts, and (III) the context-aware prompter, which captures contextual knowledge from relevant instances and generates dynamic prompts to largely enhance the MMT’s robustness. Extensive experiments conducted on three real-world datasets show that RAGPT consistently outperforms all competitive baselines in handling incomplete modality problems.

Face retouching aims to remove facial imperfections from image and videos while at the same time preserving face attributes. The existing methods are designed to perform non-interactive end-to-end retouching, while the ability to interact with users is highly demanded in downstream applications. In this paper, we propose RetouchGPT, a novel framework that leverages Large Language Models (LLMs) to guide the interactive retouching process. Towards this end, we design an instruction-driven imperfection prediction module to accurately identify imperfections by integrating textual and visual features. To learn imperfection prompts, we further incorporate a LLM-based embedding module to fuse multi-modal conditioning information. The prompt-based feature modification is performed in each transformer block, such that the imperfection features are suppressed and replaced with the features of normal skin progressively. Extensive experiments have been performed to verify effectiveness of our design elements and demonstrate that RetouchGPT is a useful tool for interactive face retouching and achieves superior performance over state-of-the-arts.

A few recent studies have shown the benefits of using centrally pre-trained models to initialize federated learning (FL). However, existing methods do not generalize well when faced with an arbitrary set of downstream FL tasks. Specifically, they often (i) achieve limited accuracy, especially with unseen downstream labels, and (ii) result in significant accuracy variance, failing to provide a balanced performance across clients. To address these challenges, we propose CoPreFL, a collaborative/distributed pre-training approach that robustly initializes for downstream FL tasks. CoPreFL leverages model-agnostic meta-learning (MAML) that tailors the global model to mimic heterogeneous and unseen FL scenarios, resulting in a pre-trained model that is rapidly adaptable to any FL task. Our MAML procedure integrates performance variance into the meta-objective function, balancing performance across clients rather than solely optimizing for accuracy. Extensive experiments show that CoPreFL significantly enhances average accuracy and reduces variance in arbitrary downstream FL tasks with unseen/seen labels, outperforming various pre-training baselines. Additionally, CoPreFL proves compatible with different well-known FL algorithms used in downstream tasks, boosting performance in each case.

Knowledge tracing (KT) models students' knowledge states and predicts their future performance based on their historical interaction data. However, attention based KT models struggle to accurately capture diverse forgetting behaviors in ever-growing interaction sequences. First, existing models use uniform time decay matrices, conflating forgetting representations with problem relevance. Second, the fixed-length window prediction paradigm fails to model continuous forgetting processes in expanding sequences. To address these challenges, this paper introduces LefoKT, a unified architecture that enhances attention based KT models by incorporating proposed relative forgetting attention. LefoKT improves forgetting modeling through relative forgetting attention to decouple forgetting patterns from problem relevance. It also enhances attention based KT models' length extrapolation capability for capturing continuous forgetting processes in ever-growing interaction sequences. Extensive experimental results on three datasets validate the effectiveness of LefoKT.

U-Net is a widely used model for medical image segmentation, renowned for its strong feature extraction capabilities and U-shaped design, which incorporates skip connections to preserve critical information. However, its decoders exhibit information-specific preferences for the supplementary content provided by skip connections, instead of adhering to a strict one-to-one correspondence, which limits its flexibility across diverse tasks. To address this limitation, we propose the Task-Adaptive Mixture of Skip Connections (TA-MoSC) module, inspired by the Mixture of Experts (MoE) framework. TA-MoSC innovatively reinterprets skip connections as a task allocation problem, employing a routing mechanism to adaptively select expert combinations at different decoding stages. By introducing MoE, our approach enhances the sparsity of the model, and lightweight convolutional experts are shared across all skip connection stages, with a Balanced Expert Utilization (BEU) strategy ensuring that all experts are effectively trained, maintaining training balance and preserving computational efficiency. Our approach introduces minimal additional parameters to the original U-Net but significantly enhances its performance and stability. Experiments on GlaS, MoNuSeg, Synapse, and ISIC16 datasets demonstrate state-of-the-art accuracy and better generalization across diverse tasks. Moreover, while this work focuses on medical image segmentation, the proposed method can be seamlessly extended to other segmentation tasks, offering a flexible and efficient solution for diverse applications.

Blind-spot networks (BSN) have been prevalent neural architectures in self-supervised image denoising (SSID). However, most existing BSNs are conducted with convolution layers. Although transformers have shown the potential to overcome the limitations of convolutions in many image restoration tasks, the attention mechanisms may violate the blind-spot requirement, thereby restricting their applicability in BSN. To this end, we propose to analyze and redesign the channel and spatial attentions to meet the blind-spot requirement. Specifically, channel self-attention may leak the blind-spot information in multi-scale architectures, since the downsampling shuffles the spatial feature into channel dimensions. To alleviate this problem, we divide the channel into several groups and perform channel attention separately. For spatial self-attention, we apply an elaborate mask to the attention matrix to restrict and mimic the receptive field of dilated convolution. Based on the redesigned channel and window attentions, we build a Transformer-based Blind-Spot Network (TBSN), which shows strong local fitting and global perspective abilities. Furthermore, we introduce a knowledge distillation strategy that distills TBSN into smaller denoisers to improve computational efficiency while maintaining performance. Extensive experiments on real-world image denoising datasets show that TBSN largely extends the receptive field and exhibits favorable performance against state-of-the-art SSID methods.

Pseudo-label learning methods have been widely applied in weakly-supervised temporal action localization. Existing works directly utilize weakly-supervised base model to generate instance-level pseudo-labels for training the fully-supervised detection head. We argue that the noise in pseudo-labels would interfere with the learning of fully-supervised detection head, leading to significant performance leakage. Issues with noisy labels include:(1) inaccurate boundary localization; (2) undetected short action clips; (3) multiple adjacent segments incorrectly detected as one segment. To target these issues, we introduce a two-stage noisy label learning strategy to harness every potential useful signal in noisy labels. First, we propose a frame-level pseudo-label generation model with a context-aware denoising algorithm to refine the boundaries. Second, we introduce an online-revised teacher-student framework with a missing instance compensation module and an ambiguous instance correction module to solve the short-action-missing and many-to-one problems. Besides, we apply a high-quality pseudo-label mining loss in our online-revised teacher-student framework to add different weights to the noisy labels to train more effectively. Our model outperforms the previous state-of-the-art method in detection accuracy and inference speed greatly upon the THUMOS14 and ActivityNet v1.2 benchmarks.

Understanding the 3D semantics of a scene is a fundamental problem for various scenarios such as embodied agents. While NeRFs and 3DGS excel at novel-view synthesis, previous methods for understanding their semantics have been limited to incomplete 3D understanding: their segmentation results are rendered as 2D masks that do not represent the entire 3D space. To address this limitation, we redefine the problem to segment the 3D volume and propose the following methods for better 3D understanding. We directly supervise the 3D points to train the language embedding field, unlike previous methods that anchor supervision at 2D pixels. We transfer the learned language field to 3DGS, achieving the first real-time rendering speed without sacrificing training time or accuracy. Lastly, we introduce a 3D querying and evaluation protocol for assessing the reconstructed geometry and semantics together. Code, checkpoints, and annotations are available at the project page.

In 3D human action recognition, limited supervised data makes it challenging to fully tap into the modeling potential of powerful networks such as transformers. As a result, researchers have been actively investigating effective self-supervised pre-training strategies. For example, MAMP shows that instead of following the prevalent masked joint reconstruction, explicit masked motion reconstruction is key to the success of learning effective feature representation for 3D action recognition. However, we find that if we make a simple and effective change to the reconstructed target of masked joint reconstruction, masked joint reconstruction can achieve the same results as masked motion reconstruction. The devil is in the special characteristic of 3D skeleton data and the normalization process of training targets. We need to dig for all effective information of targets during normalization. Besides, considering that mask data reconstruction focuses more on learning local relations in input data for fulfilling the reconstruction task, instead of modeling the relation among samples, we further employ contrastive learning to learn more discriminative 3D action representations. We show that contrastive learning can consistently boost the performance of model pre-trained by masked joint prediction under various settings, especially in the semi-supervised setting that has a very limited number of labeled samples. Extensive experiments on NTU-60, NTU-120, and PKU-MMD datasets show that the proposed pre-training strategy achieves state-of-the-art results without bells and whistles.

Spike cameras, as innovative neuromorphic devices, generate continuous spike streams to capture high-speed scenes with lower bandwidth and higher dynamic range than traditional RGB cameras. However, reconstructing high-quality images from the spike input under low-light conditions remains challenging. Conventional learning-based methods often rely on the synthetic dataset as the supervision for training. Still, these approaches falter when dealing with noisy spikes fired under the low-light environment, leading to further performance degradation in the real-world dataset. This phenomenon is primarily due to inadequate noise modelling and the domain gap between synthetic and real datasets, resulting in recovered images with unclear textures, excessive noise, and diminished brightness. To address these challenges, we introduce a novel spike-to-image reconstruction framework SpikeCLIP that goes beyond traditional training paradigms. Leveraging the CLIP model's powerful capability to align text and images, we incorporate the textual description of the captured scene and unpaired high-quality datasets as the supervision. Textual descriptions provide additional context that guides the network's feature reconstruction, while high-quality datasets help produce sharp latent images. Our experiments on real-world low-light datasets U-CALTECH and U-CIFAR demonstrate that SpikeCLIP significantly enhances texture details and the luminance balance of recovered images. Furthermore, the reconstructed images are well-aligned with the broader visual features needed for downstream tasks, ensuring more robust and versatile performance in challenging environments.

Graph neural networks (GNNs) have shown promise in integrating protein-protein interaction (PPI) networks for identifying cancer genes in recent studies. However, due to the insufficient modeling of the biological information in PPI networks, more faithfully depiction of complex protein interaction patterns for cancer genes within the graph structure remains largely unexplored. This study takes a pioneering step toward bridging biological anomalies in protein interactions caused by cancer genes to statistical graph anomaly. We find a unique graph anomaly exhibited by cancer genes, namely weight heterogeneity, which manifests as significantly higher variance in edge weights of cancer gene nodes within the graph. Additionally, from the spectral perspective, we demonstrate that the weight heterogeneity could lead to the "flattening out" of spectral energy, with a concentration towards the extremes of the spectrum. Building on these insights, we propose the HIerarchical-Perspective Graph Neural Network (HIPGNN) that not only determines spectral energy distribution variations on the spectral perspective, but also perceives detailed protein interaction context on the spatial perspective. Extensive experiments are conducted on two reprocessed datasets STRINGdb and CPDB, and the experimental results demonstrate the superiority of HIPGNN.

To preserve user privacy in recommender systems, federated recommendation (FR) based on federated learning (FL) emerges, keeping the personal data on the local client and updating a model collaboratively. Unlike FL, FR has a unique sparse aggregation mechanism, where the embedding of each item is updated by only partial clients, instead of full clients in a dense aggregation of general FL. Recently, as an essential principle of FL, model security has received increasing attention, especially for Byzantine attacks, where malicious clients can send arbitrary updates. The problem of exploring the Byzantine robustness of FR is particularly critical since in the domains applying FR, e.g., e-commerce, malicious clients can be injected easily by registering new accounts. However, existing Byzantine works neglect the unique sparse aggregation of FR, making them unsuitable for our problem. Thus, we make the first effort to investigate Byzantine attacks on FR from the perspective of sparse aggregation, which is non-trivial: it is not clear how to define Byzantine robustness under sparse aggregations and design Byzantine attacks under limited knowledge/capability. In this paper, we reformulate the Byzantine robustness under sparse aggregation by defining the aggregation for a single item as the smallest execution unit. Then we propose a family of effective attack strategies, named Spattack, which exploit the vulnerability in sparse aggregation and are categorized along the adversary's knowledge and capability. Extensive experimental results demonstrate that Spattack can effectively prevent convergence and even break down defenses under a few malicious clients, raising alarms for securing FR systems.

The emergence of Vision-Language Models (VLMs) is significant advancement in integrating computer vision with Large Language Models (LLMs) to enhance multi-modal machine learning capabilities. However, this progress has made VLMs vulnerable to advanced adversarial attacks, raising concerns about reliability. Objective of this paper is to assess resilience of VLMs against jailbreak attacks that can compromise model safety compliance and result in harmful outputs. To evaluate VLM's ability to maintain robustness against adversarial input perturbations, we propose novel metric called \textbf{Retention Score}. Retention Score is multi-modal evaluation metric that includes Retention-I and Retention-T scores for quantifying jailbreak risks in visual and textual components of VLMs. Our process involves generating synthetic image-text pairs using conditional diffusion model. These pairs are then predicted for toxicity score by VLM alongside toxicity judgment classifier. By calculating margin in toxicity scores, we can quantify robustness of VLM in attack-agnostic manner. Our work has four main contributions. First, we prove that Retention Score can serve as certified robustness metric. Second, we demonstrate that most VLMs with visual components are less robust against jailbreak attacks than corresponding plain VLMs. Additionally, we evaluate black-box VLM APIs and find that security settings in Google Gemini significantly affect score and robustness. Moreover, robustness of GPT4V is similar to medium settings of Gemini. Finally, our approach offers time-efficient alternative to existing adversarial attack methods and provides consistent model robustness rankings when evaluated on VLMs including MiniGPT-4, InstructBLIP, and LLaVA.

Diffusion models are prominent in image generation for producing detailed and realistic images from Gaussian noises. However, they often encounter instability issues in image restoration tasks, e.g., super-resolution. Existing methods typically rely on multiple runs to find an initial noise that produces a reasonably restored image. Unfortunately, these methods are computationally expensive and time-consuming without guaranteeing stable and consistent performance. To address these challenges, we propose a novel Predictive Noise Fusion Strategy (PNFS) that predicts pixel-wise errors in the restored image and combines different noises to generate a more effective noise. Extensive experiments show that PNFS significantly improves the stability and performance of diffusion models in super-resolution, both quantitatively and qualitatively. Furthermore, PNFS can be flexibly integrated into various diffusion models to enhance their stability.

The metro flow in Urban Rail Transit Systems (URTS) differs from other urban traffic flows because it is characterized by: (1) highly predetermined scheduling; and (2) interactively dynamic dependencies over the fixed physical infrastructure that vary with spatiotemporal and environmental factors. Notwithstanding the advances in graph neural networks, existing efforts fail to fully capture the characteristics and complex spatiotemporal dynamics specific to metro flow, as the innate graph-aware interactions underlying a metro flow are frequently affected by an amalgamation of: intrinsic connectivity, environmental associations, and flow-activated correlation, which usually dynamically evolve over time while containing redundant signals. We propose ReDyNet, a novel Responsive Dynamic Graph Neural Network to accurately understand the spatiotemporal dynamics of metro flow and external factors. Specifically, it employs a responsive mechanism that adapts to variations in metro flow and external influences, ensuring the construction of an appropriate dynamic graph. In addition, ReDyNet follows the merits of information bottleneck (IB) theory with redundancy disentanglement to enhance the clarity and precision of contextual spatial signals. Our experiments conducted on three real-world metro passenger flow datasets demonstrate that the proposed ReDyNet outperforms several representative baselines.

Responsibility plays a key role in the development and deployment of trustworthy autonomous systems. In this paper, we focus on the problem of strategic reasoning in probabilistic multi-agent systems with responsibility-aware agents. We introduce the logic PATL+R, a variant of Probabilistic Alternating-time Temporal Logic. The novelty of PATL+R lies in its incorporation of modalities for causal responsibility, providing a framework for responsibility-aware multi-agent strategic reasoning. We present an approach to synthesise joint strategies that satisfy an outcome specified in PATL+R, while optimising the share of expected causal responsibility and reward. This provides a notion of balanced distribution of responsibility and reward gain among agents. To this end, we utilise the Nash equilibrium as the solution concept for our strategic reasoning problem and demonstrate how to compute responsibility-aware Nash equilibrium strategies via a reduction to parametric model checking of concurrent stochastic multi-player games.

The classic Resource Constrained Shortest Path (RCSP) problem aims to find a cost optimal path between a pair of nodes in a network such that the resources used in the path are within a given limit. Having been studied for over a decade, RCSP has seen recent solutions that utilize heuristic-guided search to solve the constrained problem faster. Building upon the bidirectional A* search paradigm, this paper introduces a novel constrained search framework that uses efficient pruning strategies to allow for accelerated and effective RCSP search in large-scale networks. Results show that, compared to the state of the art, our enhanced framework can significantly reduce the constrained search time, achieving speed-ups of over to two orders of magnitude.

Personalized text-to-image generation methods can generate customized images based on the reference images, which have garnered wide research interest. Recent methods propose a finetuning-free approach with a decoupled cross-attention mechanism to generate personalized images requiring no test-time finetuning. However, when multiple reference images are provided, the current decoupled cross-attention mechanism encounters the object confusion problem and fails to map each reference image to its corresponding object, thereby seriously limiting its scope of application. To address the object confusion problem, in this work we investigate the relevance of different positions of the latent image features to the target object in diffusion model, and accordingly propose a weighted-merge method to merge multiple reference image features into the corresponding objects. Next, we integrate this weighted-merge method into existing pre-trained models and continue to train the model on a multi-object dataset constructed from the open-sourced SA-1B dataset. To mitigate object confusion and reduce training costs, we propose an object quality score to estimate the image quality for the selection of high-quality training samples. Furthermore, our weighted-merge training framework can be employed on single-object generation when a single object has multiple reference images. The experiments verify that our method achieves superior performance to the state-of-the-arts on the Concept101 dataset and DreamBooth dataset of multi-object personalized image generation, and remarkably improves the performance on single-object personalized image generation.

Defocus deblurring is a challenging task due to the spatially varying nature of defocus blur with multiple plausible solutions of a single given image. However, most existing methods falter when faced with extensive and variable defocus blur, either ignoring it or relying on additional loss functions to enhance perceptual quality. This often results in unrealistic reconstructions and compromised generalizability. In this paper, we propose a novel Residual Diffusion Deblurring Model framework for single image defocus deblurring. Our approach integrates a pre-trained defocus map estimator and a lightweight pre-deblur module with a learnable receptive field, providing crucial posterior information to effectively address large-scale and varying shaped defocus blur. In addition, a carefully-design denoising network enables the generation of diverse reconstructions from a single input. This approach not only significantly improves the perceptual quality of defocus deblurring outputs through multi-step residual learning, but also offers a more efficient inference strategy. Experimental results demonstrate that our method achieves competitive performance on real-world defocus deblurring image datasets across both perceptual and distortion evaluation metrics.