AAAI 2025 Papers

Decision-focused learning (DFL) offers an end-to-end approach to the predict-then-optimize (PO) framework by training predictive models directly on decision loss (DL), enhancing decision-making performance within PO contexts. However, the implementation of DFL poses distinct challenges. Primarily, DL can result in deviation from the physical significance of the predictions under limited data. Additionally, some predictive models are non-differentiable or black-box, which cannot be adjusted using gradient-based methods. To tackle the above challenges, we propose a novel framework, Decision-Focused Fine-tuning (DFF), which embeds the DFL module into the PO pipeline via a novel bias correction module. DFF is formulated as a constrained optimization problem that maintains the proximity of the DL-enhanced model to the original predictive model within a defined trust region. We theoretically prove that DFF strictly confines prediction bias within a predetermined upper bound, even with limited datasets, thereby substantially reducing prediction shifts caused by DL under limited data. Furthermore, the bias correction module can be integrated into diverse predictive models, enhancing adaptability to a broad range of PO tasks. Extensive evaluations on synthetic and real-world datasets, including network flow, portfolio optimization, and resource allocation problems with different predictive models, demonstrate that DFF not only improves decision performance but also adheres to fine-tuning constraints, showcasing robust adaptability across various scenarios.

Videos inherently contain complex temporal dynamics across various spatial directions, often entangled in ways that obscure effective dynamic extraction. Previous studies typically process video spatiotemporal features without disentangling, which hampers their ability to extract dynamic information. Additionally, the extraction of dynamics is disrupted by transient high-dynamic information in video sequences, e.g., noise or flicker, which has received limited attention in the literature. To tackle those problems, this paper proposes the Disentangling and Filtering Dynamics Network (DFDNet). Firstly, to disentangle the interwoven dynamics, DFDNet decomposes the spatially encoded video sequences into lower dimensional sequences. Secondly, a learnable threshold filter is proposed to eliminate the transient high-dynamic information. Thirdly, the model incorporates an MLP to extract the temporal dependencies from the disentangled and filtered sequences. DFDNet demonstrates competitive performance across four chosen datasets, including both low and high-resolution videos. Specifically, on the low-resolution Moving MNIST dataset, DFDNet achieves a 19% improvement on MSE over the previous state-of-the-art model. On the high-resolution SJTU4K dataset, it outperforms the previous state-of-the-art model by 10% on the LPIPS metric under similar inference time.

Membership Inference Attack (MIA) aims to determine if a specific sample is present in the training dataset of a target machine learning model. Previous MIAs against fine-tuned Large Language Models (LLMs) either fail to address the unique challenges in the fine-tuned setting or rely on strong assumption of the training data distribution. This paper proposes a distribution-free MIA framework tailored for fine-tuned LLMs, named DF-MIA. We recognize that samples await to test can serve as a valuable reference dataset for fine-tuning reference models. By enhancing the signals of non-member samples within this reference dataset, we can achieve a more reliable and practical calibration of probabilities, improving the differentiation between members and non-members. Leveraging these insights, we have developed a two-stage framework that employs specially designed data augmentation and perturbation techniques to prioritize the significance of non-members and mitigate the influence of potential members within the reference dataset. We evaluate our method on three representative LLM models ranging from 1B to 8B on three datasets. The results demonstrate that the DF-MIA significantly enhances the performance of MIA.

Multimodal aspect-based sentiment analysis (MABSA) integrates text and images to perform fine-grained sentiment analysis on specific aspects, enhancing the understanding of user opinions in various applications. Existing methods use modality alignment for information interaction and fusion between images and text, but an inherent gap between these two modalities necessitates a more direct bridging mechanism to effectively connect image understanding with text content. For this, we propose the Descriptions Enhanced Question-Answering Framework (DEQA), which generates descriptions of images using GPT-4, leveraging the multimodal large language model to provide more direct semantic context of images. In DEQA, to help the model better understand the task's purpose, we frame MABSA as a multi-turn question-answering problem to add semantic guidance and hints. We input text, image, and description into separate experts in various combinations, allowing each expert to focus on different features and thereby improving the comprehensive utilization of input information. By integrating these expert outputs within a multi-turn question-answering format, we employ a multi-expert ensemble decision-making approach to produce the final prediction results. Experimental results on two widely-used datasets demonstrate that our method achieves state-of-the-art performance. Furthermore, our framework substantially outperforms GPT-4o and other multimodal large language models, showcasing its superior effectiveness in multimodal sentiment analysis.

Recent research demonstrates the effectiveness of using pre-trained language models for legal case retrieval. Most of the existing works focus on improving the representation ability for the contextualized embedding of the [CLS] token and calculate relevance using textual semantic similarity. However, in the legal domain, textual semantic similarity does not always imply that the cases are relevant enough. Instead, relevance in legal cases primarily depends on the similarity of key facts that impact the final judgment. Without proper treatments, the discriminative ability of learned representations could be limited since legal cases are lengthy and contain numerous non-key facts. To this end, we introduce DELTA, a discriminative model designed for legal case retrieval. The basic idea involves pinpointing key facts in legal cases and pulling the contextualized embedding of the [CLS] token closer to the key facts while pushing away from the non-key facts, which can warm up the case embedding space in an unsupervised manner. To be specific, this study brings the word alignment mechanism to the contextual masked auto-encoder. First, we leverage shallow decoders to create information bottlenecks, aiming to enhance the representation ability. Second, we employ the deep decoder to enable ``translation'' between different structures, with the goal of pinpointing key facts to enhance discriminative ability. Comprehensive experiments conducted on publicly available legal benchmarks show that our approach can outperform existing state-of-the-art methods in legal case retrieval. It provides a new perspective on the in-depth understanding and processing of legal case documents.

Speech-driven 3D facial animation has garnered lots of attention thanks to its broad range of applications. Despite recent advancements in achieving realistic lip motion, current methods fail to capture the nuanced emotional undertones conveyed through speech and produce monotonous facial motion. These limitations result in blunt and repetitive facial animations, reducing user engagement and hindering their applicability. To address these challenges, we introduce DEEPTalk, a novel approach that generates diverse and emotionally rich 3D facial expressions directly from speech inputs. To achieve this, we first train DEE (Dynamic Emotion Embedding), which employs probabilistic contrastive learning to forge a joint emotion embedding space for both speech and facial motion. This probabilistic framework captures the uncertainty in interpreting emotions from speech and facial motion, enabling the derivation of emotion vectors from its multifaceted space. Moreover, to generate dynamic facial motion, we design TH-VQVAE (Temporally Hierarchical VQ-VAE) as an expressive and robust motion prior overcoming limitations of VAEs and VQ-VAEs. Utilizing these strong priors, we develop DEEPTalk, a talking head generator that non-autoregressively predicts codebook indices to create dynamic facial motion, incorporating a novel emotion consistency loss. Extensive experiments on various datasets demonstrate the effectiveness of our approach in creating diverse, emotionally expressive talking faces that maintain accurate lip-sync. Our project page is available at https://whwjdqls.github.io/deeptalk.github.io/.

Alzheimer’s Disease (AD) is an irreversible neurodegenerative disease affecting 50 million people worldwide. Low-cost, accurate identification of key markers of AD is crucial for timely diagnosis and intervention. Language impairment is one of the earliest signs of cognitive decline, which can be used to discriminate AD patients from normal control (NC) individuals. Patient-interviewer dialogues may be used to detect such impairments, but they are often mixed with ambiguous, noisy, and irrelevant information, making the AD detection task difficult. Moreover, the limited availability of AD speech samples and variability in their speech styles pose significant challenges in developing robust speech-based AD detection models. To address these challenges, we propose DECT, a novel speech-based domain-specific approach leveraging large language models (LLMs) for fine-grained linguistic analysis and label-switched label-preserved (LSLP) data generation. Our study presents four novelties: (1) We harness the summarizing capabilities of LLMs to identify and distill key Cognitive-Linguistic (CL) information (atoms) from noisy speech transcripts, effectively filtering irrelevant information. (2) We leverage the inherent linguistic knowledge of LLMs to extract linguistic markers from unstructured and heterogeneous audio transcripts. (3) We exploit the compositional ability of LLMs to generate LSLP AD speech transcripts consisting of diverse linguistic patterns to overcome the speech data scarcity challenge and enhance the robustness of AD detection models. (4) We use the augmented AD textual speech transcript dataset and a more fine-grained representation of AD textual speech transcript data to fine-tune the AD detection model. The results have shown that DECT, an integrated, LLM-assisted, speech-based AD detection model demonstrates superior model performance with an 11% improvement in AD detection accuracy on the datasets from DementiaBank compared to the baselines.

Image change captioning (ICC) poses great challenges stemming from describing subtle differences between two similar images in natural language, significantly increasing the complexity of feature extraction and cross-modal learning compared to the image captioning task. Existing ICC methods often suffer from two key challenges: 1) Massive irrelevant information of uni-image features leads to suboptimal visual difference representations; 2) Imprecise inter-modality correspondence degrades the quality of generated captions. This paper proposes a Difference-aware Contrastive Diffusion Model with Adversarial Perturbations (DECIDER) for ICC due to the excellent performance of diffusion models in image/text generation. Technically, difference-aware cross-modal learning is developed to suppress irrelevant information and learn compact yet robust visual difference representations. This is achieved by optimizing a novel objective mathematically derived from the information bottleneck principle that excels in filtering redundant features and highlighting differences. Furthermore, we propose to dynamically generate ``hard'' positive and negative samples via adversarial perturbations, which are involved in contrastive diffusion training with a tighter variational bound. This design encourages our DECIDER to excavate and construct complex correspondences between visual differences and captions, thereby improving generalization performance. Extensive experiments on four datasets demonstrate that DECIDER significantly exceeds state-of-the-art performance.

Most existing just noticeable difference (JND) methods primarily integrate specific masking effects in a single domain. However, these single-domain JND methods struggle with the structural discrepancies in multi-source content images, limiting their effectiveness in visual redundancy estimation. To address this issue, we propose a dual domain encoder that combines spatial and frequency features to comprehensively capture visual patterns. Our design includes spatial pattern balance and frequency detail correction modules to balance global and local patterns and correct low- and high-frequency distributions. Additionally, we develop a dual domain decoder to effectively extract multi-scale pattern redundancies and integrate them with detail redundancies in the frequency domain. Experiments demonstrate the effectiveness and robustness of our proposed method in handling structural discrepancies in multi-source content images.

Despite the advanced long-sequence modeling of Mamba, which has expanded its applications in image restoration, there remains a lack of exploration combining its strengths with the specific characteristics of JPEG image restoration, where high-frequency components are lost after the Discrete Cosine Transform (DCT). To address this, we introduce DCTMamba, a new framework designed to apply Mamba more effectively to JPEG image restoration. Specifically, our method integrates the Discrete Cosine Transform (DCT) into the Mamba to establish the sequential scanning from lower to higher frequencies, enabling the network to initially reconstruct coarse structures and progressively refine the image with more intricate details. Furthermore, recognizing the variable frequency distributions that arise from DCT transformations across different image sizes, we have developed Scale-Adaptive Normalization to manage these variations adeptly. Comprehensive experiments confirm that DCTMamba significantly outperforms existing solutions, achieving high fidelity in both coarse structures and fine details.CTMamba significantly outperforms existing solutions, achieving high fidelity in both coarse structures and fine details.

Knowledge distillation (KD) has recently gained great success in the field of object detection. By transferring the knowledge of the spatial or channel domain from the teacher model to the student model, it allows for a more compact representation with minimal performance loss. Despite this progress, existing KD methods typically treat knowledge from spatial or channel domains independently, ignoring the exploitation of the mutual relationship between these domains. In this work, we first explore the connection between spatial and channel domains and find there exists a strong correlation between them, i.e. the salient channels tend to contain significant object regions in the spatial domain. Motivated by this observation, we propose DCSF-KD, a novel Dynamic Channel-wise Spatial Feature Knowledge Distillation framework for object detection by fully exploiting both spatial and channel knowledge. Specifically, we introduce channel-wise spatial feature distillation and global channel attention distillation, using information from both domains to improve the accuracy of the student network. Experiments demonstrate that our DCSF-KD outperforms existing detection methods on both homogeneous and heterogeneous teacher-student network pairs. For example, when using the MaskRCNN-Swin detector as the teacher, and based on RetinaNet and FCOS with ResNet-50 on MS COCO, our DCSF-KD can achieve 41.9% and 44.1% mAP, respectively.

Causal learning tackles the computationally demanding task of estimating causal graphs. This paper introduces a new divide-and-conquer approach for causal graph learning, called DCILP. In the divide phase, the Markov blanket MB(Xi) of each variable Xi is identified, and causal learning subproblems associated with each MB(Xi) are independently addressed in parallel. This approach benefits from a more favorable ratio between the number of data samples and the number of variables considered. In counterpart, it can be adversely affected by the presence of hidden confounders, as variables external to MB(Xi) might influence those within it. The reconciliation of the local causal graphs generated during the divide phase is a challenging combinatorial optimization problem, especially in large-scale applications. The main novelty of DCILP is an original formulation of this reconciliation as an integer linear programming (ILP) problem, which can be delegated and efficiently handled by an ILP solver. Through experiments on medium to large scale graphs, and comparisons with state-of-the-art methods, DCILP demonstrates significant improvements in terms of computational complexity, while preserving the learning accuracy on real-world problem and suffering at most a slight loss of accuracy on synthetic problems.

Solutions to time-varying problems are crucial for research areas such as predicting changes in human body shape over time. While recurrent neural networks have made significant advancements in this field, their reliance on centralized processing has led to challenges such as model silos and data isolation. In response, distributed AI systems like federated learning have emerged to facilitate dynamic collaboration among models; however, they still depend on central coordinators, which pose risks to system security and efficiency. Moreover, traditional federated learning primarily supports homogeneous models and lacks effective strategies for the interaction of heterogeneous models. To address these limitations, we propose a novel method called Dynamic Collaboration of Heterogeneous Models (DCHM), based on Isomerism Learning, which leverages a consortium blockchain network to enhance model credibility and facilitate coordination among heterogeneous models. Additionally, we introduce a Distributed Hierarchical Aggregation (DHA) algorithm that enables permissioned nodes within each group to aggregate local model results and share them for standardized processing. After several iterative cycles, these nodes perform secondary integration of local results to produce global outcomes. Experimental results demonstrate that DCHM effectively analyzes the temporal variability of body shape changes with high efficiency.

Index tracking is a popular passive investment strategy aimed at optimizing portfolios, but fully replicating an index can lead to high transaction costs. To address this, partial replication have been proposed. However, the cardinality constraint renders the problem non-convex, non-differentiable, and often NP-hard, leading to the use of heuristic or neural network-based methods, which can be non-interpretable or have NP-hard complexity. To overcome these limitations, We propose a Differentiable Cardinality Constraint (DCC) for index tracking and introduce a floating-point precision-aware method to address implementation issues. We theoretically prove our methods calculate cardinality accurately and enforce actual cardinality with polynomial time complexity. We propose the range of the hyperparameter ensures that our method has no error in real implementations, based on theoretical proof and experiment. Our method applied to mathematical method outperforms baseline methods across various datasets, demonstrating the effectiveness of the identified hyperparameter.

Incremental object detection (IOD) aims to cultivate an object detector that can continuously localize and recognize novel classes while preserving its performance on previous classes. Existing methods achieve certain success by improving knowledge distillation and exemplar replay for transformer-based detection frameworks, but the intrinsic forgetting mechanisms remain underexplored. In this paper, we dive into the cause of forgetting and discover forgetting imbalance between localization and recognition in transformer-based IOD, which means that localization is less-forgetting and can generalize to future classes, whereas catastrophic forgetting occurs primarily on recognition. Based on these insights, we propose a Divide-and-Conquer Amnesia (DCA) strategy, which redesigns the transformer-based IOD into a localization-then-recognition process. DCA can well maintain and transfer the localization ability, leaving decoupled fragile recognition to be specially conquered. To reduce feature drift in recognition, we leverage semantic knowledge encoded in pre-trained language models to anchor class representations within a unified feature space across incremental tasks. This involves designing a duplex classifier fusion and embedding class semantic features into the recognition decoding process in the form of queries. Extensive experiments validate that our approach achieves state-of-the-art performance, especially for long-term incremental scenarios. For example, under the four-step setting on MS-COCO, our DCA strategy significantly improves the final AP by 6.9%.

Point cloud completion aims to reconstruct complete 3D shapes from partial 3D point clouds. With advancements in deep learning techniques, various methods for point cloud completion have been developed. Despite achieving encouraging results, a significant issue remains: these methods often overlook the variability in point clouds sampled from a single 3D object surface. This variability can lead to ambiguity and hinder the achievement of more precise completion results. Therefore, in this study, we introduce a novel point cloud completion network, namely Dual-Codebook Point Completion Network (DC-PCN), following an encder-decoder pipeline. The primary objective of DC-PCN is to formulate a singular representation of sampled point clouds originating from the same 3D surface. DC-PCN introduces a dual-codebook design to quantize point-cloud representations from a multilevel perspective. It consists of an encoder-codebook and a decoder-codebook, designed to capture distinct point cloud patterns at shallow and deep levels. Additionally, to enhance the information flow between these two codebooks, we devise an information exchange mechanism. This approach ensures that crucial features and patterns from both shallow and deep levels are effectively utilized for completion. Extensive experiments on the PCN, ShapeNet_Part, and ShapeNet34 datasets demonstrate the state-of-the-art performance of our method.

Lifelong person re-identification (LReID) is an important but challenging task that suffers from catastrophic forgetting due to significant domain gaps between training steps. Existing LReID approaches typically rely on data replay and knowledge distillation to mitigate this issue. However, data replay methods compromise data privacy by storing historical exemplars, while knowledge distillation methods suffer from limited performance due to the cumulative forgetting of undistilled knowledge. To overcome these challenges, we propose a novel paradigm that models and rehearses the distribution of the old domains to enhance knowledge consolidation during the new data learning, possessing a strong anti-forgetting capacity without storing any exemplars. Specifically, we introduce an exemplar-free LReID method called Distribution Rehearsing via Adaptive Style Kernel Learning (DASK). DASK includes a Distribution Rehearser Learning mechanism that learns to transform arbitrary distribution data into the current data style at each learning step. To enhance the style transfer capacity, an Adaptive Kernel Prediction network is explored to achieve an instance-specific distribution adjustment. Additionally, we design a Distribution Rehearsing-driven LReID Training module, which rehearses old distribution based on the new data via the old AKPNet model, achieving effective knowledge accumulation. Experimental results show our DASK outperforms the existing methods by 3.6%-6.8% and 4.5%-6.5% on seen and unseen domains, respectively.

Abstract visual reasoning (AVR) is a critical ability of humans, and it has been widely studied, but arithmetic visual reasoning, a unique task in AVR to reason over number sense, is less studied in the literature. To facilitate this research, we construct a Machine Number Reasoning (MNR) dataset to assess the model's ability in arithmetic visual reasoning over number sense and spatial layouts. To solve the MNR tasks, we propose a Dual-branch Arithmetic Regression Reasoning (DARR) framework, which includes an Intra-Image Arithmetic Regression Reasoning (IIARR) module and a Cross-Image Arithmetic Regression Reasoning (CIARR) module. The IIARR includes a set of Intra-Image Regression Blocks to identify the correct number orders and the underlying arithmetic rules within individual images, and an Order Gate to determine the correct number order. The CIARR establishes the arithmetic relations across different images through a `3-to-1' regressor and a set of `2-to-1' regressors, with a Selection Gate to select the most suitable `2-to-1' regressor and a gated fusion to combine the two kinds of regressors. Experiments on the MNR dataset show that the DARR outperforms state-of-the-art models for arithmetic visual reasoning.

Point Transformers (PoinTr) have shown great potential in point cloud completion recently. Nevertheless, effective domain adaptation that improves transferability toward target domains remains unexplored. In this paper, we delve into this topic and empirically discover that direct feature alignment on point Transformer’s CNN backbone only brings limited improvements since it cannot guarantee sequence-wise domain-invariant features in the Transformer. To this end, we propose a pioneering Domain Adaptive Point Transformer (DAPoinTr) framework for point cloud completion. DAPoinTr consists of three novel components: Domain Query-based Feature Alignment (DQFA), Point Token-wise Feature alignment (PTFA), and Voted Prediction Consistency (VPC). In particular, DQFA is presented to narrow the global domain gaps from the sequence via the presented domain proxy and domain query at the Transformer encoder and decoder, respectively. PTFA is proposed to close the local domain shifts by aligning the tokens, i.e., point proxy and dynamic query, at the Transformer encoder and decoder, respectively. VPC is designed to consider different Transformer decoders as multiple of experts (MoE) for ensembled prediction voting and pseudo-label generation. Extensive experiments with visualization on several challenging domain adaptation benchmarks demonstrate the effectiveness and superiority of our DAPoinTr compared with other state-of-the-art methods.