AAAI 2025 论文列表

To address the limitations of current Large-scale Video-Language Models (LVLMs) in fine-grained understanding and long-term temporal memory, we propose a novel video understanding approach that integrates a Vision Language Model (VLM) and a Large Language Model (LLM) with a textual memory mechanism to ensure continuity and contextual coherence. In addition, we introduce a novel evaluation metric, VAD-Score (Video Automated Description Score), to assess precision, recall, and F1 scores for events, subjects, and objects. Our approach delivers competitive results on a diverse set of videos from the DREAM-1K dataset, spanning categories such as live-action, animation, shorts, stock, and YouTube, with a focus on fine-grained comprehension.

Glass surfaces are becoming increasingly ubiquitous as modern buildings tend to use a lot of glass panels. This, however, poses substantial challenges to the operations of autonomous systems such as robots, self-driving cars, and drones, as the glass panels can become transparent obstacles to navigation. Existing works attempt to exploit various cues, including glass boundary context or reflections, as a prior. However, they are all based on input RGB images. We observe that the transmission of 3D depth sensor light through glass surfaces often produces blank regions in the depth maps, which can offer additional insights to complement the RGB image features for glass surface detection. In this work, we propose a large-scale RGB-D glass surface detection dataset, RGB-D GSD, for rigorous experiments and future research. It contains 3,009 images offering a wide range of real-world RGB-D glass surface categories, paired with precise annotations. Moreover, we propose a novel glass surface detection framework combining RGB and depth information, with two novel modules: a cross-modal context mining (CCM) module to adaptively learn individual and mutual context features from RGB and depth information, and a depth-missing aware attention (DAA) module to explicitly exploit spatial locations where missing depths occur to help detect the presence of glass surfaces. Experimental results show that our proposed model outperforms state-of-the-art methods.

Composed Image Retrieval (CIR) aims to retrieve target images from candidate set using a hybrid-modality query consisting of a reference image and a relative caption that describes the user intent. Recent studies attempt to utilize Vision-Language Pre-training Models (VLPMs) with various fusion strategies for addressing the task. However, these methods typically fail to simultaneously meet two key requirements of CIR: comprehensively extracting visual information and faithfully following the user intent. In this work, we propose CIR-LVLM, a novel framework that leverages the large vision-language model (LVLM) as the powerful user intent-aware encoder to better meet these requirements. Our motivation is to explore the advanced reasoning and instruction-following capabilities of LVLM for accurately understanding and responding the user intent. Furthermore, we design a novel hybrid intent instruction module to provide explicit intent guidance at two levels: (1) The task prompt clarifies the task requirement and assists the model in discerning user intent at the task level. (2) The instance-specific soft prompt, which is adaptively selected from the learnable prompt pool, enables the model to better comprehend the user intent at the instance level compared to a universal prompt for all instances. CIR-LVLM achieves state-of-the-art performance across three prominent benchmarks with acceptable inference efficiency. We believe this study provides fundamental insights into CIR-related fields.

Economic growth and development require a consistent supply of energy. Energy which has mainly been supplied from fossil fuels. The impacts of these on the environment such as global warming, raised an alarm on their use. As a result other sources of energy such as wind energy are used as alternatives for electricity production. Wind energy assessment nevertheless faces barriers due to its stochastic nature. This later creates various regimes, which traditional models can't always fit thereby producing poor estimates. In this work, we aim to use Large Language Models (LLMs) to predict the wind potential in a given location. Through this approach, we aim at lifting the barrier on energy problems in developing countries by providing knowledge on the state of wind energy in given locations.

Text-attributed graphs have recently garnered significant attention due to their wide range of applications in web domains. Existing methodologies employ word embedding models for acquiring text representations as node features, which are subsequently fed into Graph Neural Networks (GNNs) for training. Recently, the advent of Large Language Models (LLMs) has introduced their powerful capabilities in information retrieval and text generation, which can greatly enhance the text attributes of graph data. Furthermore, the acquisition and labeling of extensive datasets are both costly and time-consuming endeavors. Consequently, few-shot learning has emerged as a crucial problem in the context of graph learning tasks. In order to tackle this challenge, we propose a lightweight paradigm called LLM4NG, which adopts a plug-and-play approach to establish supervision signals by leveraging LLMs for node generation. Specifically, we utilize LLMs to extract semantic information from the labels and generate samples that belong to these categories as exemplars. Subsequently, we employ an edge predictor to capture the structural information inherent in the raw dataset and integrate the newly generated samples into the original graph. This approach harnesses LLMs for enhancing class-level information and seamlessly introduces labeled nodes and edges without modifying the raw dataset, thereby facilitating the node classification task in few-shot scenarios. Extensive experiments demonstrate the outstanding performance of our proposed paradigm, particularly in low-shot scenarios. For instance, in the 1-shot setting of the ogbn-arxiv dataset, LLM4NG achieves a 76% improvement over the baseline model.

Ensuring that AI systems do what we, as humans, actually want them to do, is one of the biggest open research challenges in AI alignment and safety. My research seeks to directly address this challenge by enabling AI systems to interact with humans to learn aligned and robust behaviors. The way in which robots and other AI systems behave is often the result of optimizing a reward function. However, manually designing good reward functions is highly challenging and error prone, even for domain experts. Consider trying to write down a reward function that describes good driving behavior or how you like your bed made in the morning. While reward functions for these tasks are difficult to manually specify, human feedback in the form of demonstrations or preferences are often much easier to obtain. However, human data is often difficult to interpret, due to ambiguity and noise. Thus, it is critical that AI systems take into account epistemic uncertainty over the human's true intent. My talk will give an overview of my lab's progress along the following fundamental research areas: (1) efficiently maintaining uncertainty over human intent, (2) directly optimizing behavior to be robust to uncertainty over human intent, and (3) actively querying for additional human input to reduce uncertainty over human intent.

Deep imbalanced regression (DIR), where the target values have a highly skewed distribution and are also continuous, is an intriguing yet under-explored problem in machine learning. While recent works have already shown that incorporating various classification-based regularizers can produce enhanced outcomes, the role of classification remains elusive in DIR. Moreover, such regularizers (e.g., contrastive penalties) merely focus on learning discriminative features of data, which inevitably results in ignorance of either continuity or similarity across the data. To address these issues, we first bridge the connection between the objectives of DIR and classification from a Bayesian perspective. Consequently, this motivates us to decompose the objective of DIR into a combination of classification and regression tasks, which naturally guides us toward a divide-and-conquer manner to solve the DIR problem. Specifically, by aggregating the data at nearby labels into the same groups, we introduce an ordinal group-aware contrastive learning loss along with a multi-experts regressor to tackle the different groups of data thereby maintaining the data continuity. Meanwhile, considering the similarity between the groups, we also propose a symmetric descending soft labeling strategy to exploit the intrinsic similarity across the data, which allows classification to facilitate regression more effectively. Extensive experiments on real-world datasets also validate the effectiveness of our method.

Black-box prompt tuning has become a prevalent parameter-efficient paradigm that leverages the capabilities of large language models (LLMs) for customized applications in specific downstream tasks. In practical scenarios, downstream tasks frequently involve data distributions that are heavily imbalanced. Such imbalances tend to impair performance, causing severe performance collapse in minority classes. Conducting effective imbalanced black-box prompt tuning to mitigate the adverse effects of imbalanced data distribution on prompt performance remains a significant challenge. In this paper, we propose black-box prompt tuning with first and zeroth order gradient (BPT-FZG) for handling the imbalanced data. Specifically, BPT-FZG introduces AUC maximization as the objective for prompt tuning and equivalently formulates it as a nonconvex-concave saddle point problem to avoid the construction of sample pairs from opposite classes. Indeed, BPT-FZG optimizes the latent representation of the continuous prompt in the low-dimensional subspace with AUC loss and leverages the first and zeroth order gradients alternately to update the parameters. Furthermore, we establish the theoretical convergence guarantee for BPT-FZG under common assumptions, showing that our method can find a stationary point of the objective function. Our experiments on RoBERTa-large, GPT2-XL, and Llama3 show that BPT-FZG achieves improvement on various imbalanced datasets, emphasizing the effectiveness of our methods.

In many RL applications, ensuring an agent's actions adhere to constraints is crucial for safety. Most previous methods in Action-Constrained Reinforcement Learning (ACRL) employ a projection layer after the policy network to correct the action. However projection-based methods suffer from issues like the zero gradient problem and higher runtime due to the usage of optimization solvers. Recently methods were proposed to train generative models to learn a differentiable mapping between latent variables and feasible actions to address this issue. However, generative models require training using samples from the constrained action space, which itself is challenging. To address such limitations, first, we define a target distribution for feasible actions based on constraint violation signals, and train normalizing flows by minimizing the KL divergence between an approximated distribution over feasible actions and the target. This eliminates the need to generate feasible action samples, greatly simplifying the flow model learning. Second, we integrate the learned flow model with existing deep RL methods, which restrict it to exploring only the feasible action space. Third, we extend our approach beyond ACRL to handle state-wise constraints by learning the constraint violation signal from the environment. Empirically, our approach has significantly fewer constraint violations while achieving similar or better quality in several control tasks than previous best methods.

Recent approaches to VO have significantly improved performance by using deep networks to predict optical flow between video frames. However, existing methods still suffer from noisy and inconsistent flow matching, making it difficult to handle challenging scenarios and long-sequence estimation.To overcome these challenges, we introduce Spatio-Temporal Visual Odometry (STVO), a novel deep network architecture that effectively leverages inherent spatio-temporal cues to enhance the accuracy and consistency of multi-frame flow matching. With more accurate and consistent flow matching, STVO can achieve better pose estimation through the bundle adjustment (BA).Specifically, STVO introduces two innovative components: 1) the Temporal Propagation Module that utilizes multi-frame information to extract and propagate temporal cues across adjacent frames, maintaining temporal consistency; 2) the Spatial Activation Module that utilizes geometric priors from the depth maps to enhance spatial consistency while filtering out excessive noise and incorrect matches.Our STVO achieves state-of-the-art performance on TUM-RGBD, EuRoc MAV, ETH3D and KITTI Odometry benchmarks. Notably, it improves accuracy by 77.8% on ETH3D benchmark and 38.9% on KITTI Odometry benchmark over the previous best methods.

Ensuring street food safety in developing countries is crucial due to the high prevalence of foodborne illnesses. Traditional methods of food safety assessments face challenges such as resource constraints, logistical issues, and subjective biases influenced by surveyors personal lived experiences, particularly when interacting with local communities. For instance, a local food safety inspector may inadvertently overrate the quality of infrastructure due to prior familiarity or past purchases, thereby compromising objective assessment. This subjectivity highlights the necessity for technologies that reduce human biases and enhance the accuracy of survey data across various domains. This paper proposes a novel approach based on a combination of Computer Vision and a lightweight Visual Large Language Model (VLLM) to automate the detection and analysis of critical food safety infrastructure in street food vendor environments at a field experiment in Kolkata, India. The system utilises a three-stage object extraction pipeline from the video to identify, extract and select unique representations of critical elements such as hand-washing stations, dishwashing areas, garbage bins, and water tanks. These four infrastructure items are crucial for maintaining safe food practices, irrespective of the specific methods employed by the vendors. A VLLM then analyses the extracted representations to assess compliance with food safety standards. Notably, over half of the pipeline can be processed using a user's smartphone, significantly reducing government server workload. By leveraging this decentralised approach, the proposed system decreases the analysis cost by many orders of magnitude compared to alternatives like ChatGPT or Claude 3.5. Additionally, processing data on local government servers provides better privacy and security than cloud platforms, addressing critical ethical considerations. This automated approach significantly improves efficiency, consistency, and scalability, providing a robust solution to enhance public health outcomes in developing regions.

Prompt compression is increasingly studied for its potential to reduce computational costs and alleviate the burden on language models when processing lengthy prompts. Prior research has assessed token retention and removal by calculating information entropy. However, prompt compression encounters two significant challenges: (1) Information entropy, while widely used, may not be the optimal compression metric; and (2) The semantic significance of tokens is context-dependent, which renders independent token retention decisions inadequate. We posit that the solution to these challenges lies in the intrinsic mechanism of language models. Large language models (LLMs) exhibit robust contextual processing capabilities, with recent studies on their internal dynamics revealing that the attention mechanism plays a crucial role in modeling how LLMs leverage long contexts. Building on this insight, we introduce AttnComp, a novel approach that exploits the attention mechanism within language models to guide prompt compression. Our method employs causal cross-attention from the query to the context to evaluate the significance of each token, and we develop a graph-based algorithm to efficiently cluster tokens into semantic units, thus mitigating the issue of independent dependencies. We conduct experiments on datasets for retrieval-augmented generation and multiple long tasks involving single or multi-document QA. Our proposed method, AttnComp, outperforms previous baselines and validates the contributions of our components through analytical experiments. Compared to other methods that use a causal LM for prompt compression, our approach results in shorter latency and improved performance.

Event cameras encode visual information by generating asynchronous and sparse event streams, which hold great potential for low latency and low power consumption. Despite many successful implementations of event camera-based applications, most of them accumulate the events into frames and then utilize conventional frame-based computer vision algorithms. These frame-based methods, though typically effective, diminish the inherent advantages of the event camera's low latency and low power consumption. To solve the above problems, we propose ASGCN, which efficiently processes data on an event-by-event basis and dynamically evolves into a corresponding dynamic representation, enabling low latency and high sparsity of data representation. The sparsity computation is further improved by introducing brain-inspired spiking neural networks, resulting in low power consumption for ASGCN. Extensive and diverse experiments demonstrate the energy efficiency and low latency advantages of our processing pipeline. Especially on real-world event camera datasets, our pipeline consumes more than 10,000 times less energy and achieves similar performance compared to current frame-based methods.

Source-free unsupervised domain adaptation aims to eliminate domain shifts when data from the source domain and annotation from the target domain are not available. The multi-object detection tasks in medical image analysis are constrained by patient privacy and extremely huge annotation consumption. Hence, Source-free UDA is considered a more practical approach for eliminating the domain gap. However, relevant research that explores this topic is a dearth. In this paper, we design an Anatomy-aware Alignment Teacher-Student learning method using topological consistency based on a mean-teacher framework for Source-free UDA in multiple medical object detection named AATS, including Unsupervised Structure Refinement (USR) and Graph-aware Morphology Alignment (GMA). To match the student and teacher at the low-level and visual features, we propose the USR via an unsupervised clustering algorithm to group organs in ultrasound images. Based on USR, we obtain a graph with organ relations on the teacher branch. While in the student branch, we acquire visual features to construct graphical space and optimize the model with graph propagation. Finally, to match the student and teacher, GMA is designed to align the teacher and student based on both topology and morphology information that is derived from prior medical knowledge. Four groups of adaptation experiments were conducted on available medical datasets, and the outcomes demonstrate that our approach not only achieves state-of-the-art performance but also provides substantial advantages over existing methods.

As artificial intelligence (AI) systems become increasingly deployed across the world, they are also increasingly implicated in AI incidents – harm events to individuals and society. As a result, industry, civil society, and governments worldwide are developing best practices and regulations for monitoring and analyzing AI incidents. The AI Incident Database (AIID) is a project that catalogs AI incidents and supports further research by providing a platform to classify incidents for different operational and research-oriented goals. This study reviews the AIID’s dataset of 750+ AI incidents and two independent taxonomies applied to these incidents to identify common challenges to indexing and analyzing AI incidents. We find that certain patterns of AI incidents present structural ambiguities that challenge incident databasing and explore how epistemic uncertainty in AI incident reporting is unavoidable. We therefore report mitigations to make incident processes more robust to uncertainty related to cause, extent of harm, severity, or technical details of implicated systems. With these findings, we discuss how to develop future AI incident reporting practices.

Recently, several studies have shown that utilizing contextual information to perceive target states is crucial for object tracking. They typically capture context by incorporating multiple video frames. However, these naive frame-context methods fail to consider the importance of each patch within a reference frame, making them susceptible to noise and redundant tokens, which deteriorates tracking performance. To address this challenge, we propose a new token context-aware tracking pipeline named LMTrack, designed to automatically learn high-quality reference tokens for efficient visual tracking. Embracing the principle of Less is More, the core idea of LMTrack is to analyze the importance distribution of all reference tokens, where important tokens are collected, continually attended to, and updated. Specifically, a novel Token Context Memory module is designed to dynamically collect high-quality spatio-temporal information of a target in an autoregressive manner, eliminating redundant background tokens from the reference frames. Furthermore, an effective Unidirectional Token Attention mechanism is designed to establish dependencies between reference tokens and search frame, enabling robust cross-frame association and target localization. Extensive experiments demonstrate the superiority of our tracker, achieving state-of-the-art results on tracking benchmarks such as GOT-10K, TrackingNet, and LaSOT.

Automated Program Repair (APR) is a task to automatically generate patches for the buggy code. However, most research focuses on generating correct patches while ignoring the consistency between the fixed code and the original buggy code. How to conduct adaptive bug fixing and generate patches with minimal modifications have seldom been investigated. To bridge this gap, we first introduce a novel task, namely AdaPR (Adaptive Program Repair). We then propose a two-stage approach AdaPatcher (Adaptive Patch Generator) to enhance program repair while maintaining the consistency. In the first stage, we utilize a Bug Locator with self-debug learning to accurately pinpoint bug locations. In the second stage, we train a Program Modifier to ensure consistency between the post-modified fixed code and the pre-modified buggy code. The Program Modifier is enhanced with a location-aware repair learning strategy to generate patches based on identified buggy lines, a hybrid training strategy for selective reference and an adaptive preference learning to prioritize fewer changes. The experimental results show that our approach outperforms a set of baselines by a large margin, validating the effectiveness of our two-stage framework for the newly proposed AdaPR task.

The learnware paradigm aims to establish a learnware dock system of numerous well-trained machine learning models, enabling users to reuse existing helpful models for their tasks instead of starting from scratch. Each learnware in the system is a well-established model submitted by its developer, associated with a specification generated by the learnware dock system. The specification characterizes the specialty of the corresponding model, enabling it to be identified accurately for new task requirements. Existing specification generation methods are mostly based on the Reduced Kernel Mean Embedding (RKME) technique, which uses the Maximum Mean Discrepancy (MMD) in the Reproducing Kernel Hilbert Space (RKHS) to seek a reduced set that characterizes the model's capabilities. However, existing RKME-based methods mainly utilize feature information to generate specifications by assuming the existence of the ground-truth labeling function, while leaving the label information, which is capable of providing rich semantic characterization, untouched. Furthermore, the quality of the generated specifications heavily relies on the choice of the kernels, which makes it prohibitive to adapt to all real-world scenarios. In this paper, to overcome the above limitations, we propose a novel specification approach named LANE, i.e., Label-Aware Neural Embedding. In LANE, the neural embedding space is utilized to replace the RKHS, effectively circumventing the step of kernel selection and thereby addressing the dependency on kernels in existing RKME-based specification methods. More importantly, LANE uses the label information as additional supervision to enhance the generation process, resulting in specifications of superior quality. Extensive experiments demonstrate the effectiveness and superiority of the proposed LANE approach in the learnware paradigm.

Learning with Noisy Labels (LNL) aims to improve the model generalization when facing data with noisy labels, and existing methods generally assume that noisy labels come from known classes, called closed-set noise. However, in real-world scenarios, noisy labels from similar unknown classes, i.e., open-set noise, may occur during the training and inference stage. Such open-world noisy labels may significantly impact the performance of LNL methods. In this study, we propose a novel dual-space joint learning method to robustly handle the open-world noise. To mitigate model overfitting on closed-set and open-set noises, a dual representation space is constructed by two networks. One is a projection network that learns shared representations in the prototype space, while the other is a One-Vs-All (OVA) network that makes predictions using unique semantic representations in the class-independent space. Then, bi-level contrastive learning and consistency regularization are introduced in two spaces to enhance the detection capability for data with unknown classes. To benefit from the memorization effects across different types of samples, class-independent margin criteria are designed for sample identification, which selects clean samples, weights closed-set noise, and filters open-set noise effectively. Extensive experiments demonstrate that our method outperforms the state-of-the-art methods and achieves an average accuracy improvement of 4.55\% and an AUROC improvement of 6.17\% on CIFAR80N.