AAAI 2025 Papers

Actor-Critic (AC) algorithms like SAC and TD3 were shown to perform well in a variety of continuous-action tasks. However, the theoretical basis for the pessimistic objectives these algorithms employ remains unestablished, raising questions about the specific class of policies they are implementing. In this work, we apply the expected utility hypothesis, a fundamental concept in economics, to illustrate that both pessimistic and non-pessimistic RL objectives can be interpreted through expected utility maximization using an exponential utility function. This approach reveals that pessimistic policies effectively maximize value certainty equivalent, aligning them with the optimization of risk-aware objectives. Furthermore, we propose Decoupled Policy Actor-Critic (DAC). DAC is a model-free algorithm that features two distinct actor networks: a pessimistic actor for temporal-difference learning and an optimistic actor for exploration. Our evaluations of DAC across various locomotion and manipulation tasks demonstrate improvements in sample efficiency and final performance. Remarkably, DAC, while requiring significantly fewer computational resources, matches the performance of leading model-based methods in the complex dog and humanoid domains.

Incremental few-shot semantic segmentation (IFSS) expands segmentation capacity of the trained model to segment new-class images with few samples. However, semantic meanings may shift from background to object class or vice versa during incremental learning. Moreover, new-class samples often lack representative attribute features when the new class greatly differs from the pre-learned old class. In this paper, we propose a causal framework to discuss the cause of semantic shift and incompleteness in IFSS, and we deconfound the revealed causal effects from two aspects. First, we propose a Causal Intervention Module (CIM) to resist semantic shift. CIM progressively and adaptively updates prototypes of old class, and removes the confounder in an intervention manner. Second, a Prototype Refinement Module (PRM) is proposed to complete the missing semantics. In PRM, knowledge gained from the episode learning scheme assists in fusing features of new-class and old-class prototypes. Experiments on both PASCAL-VOC 2012 and ADE20k benchmarks demonstrate the outstanding performance of our method.

Wearable Human Activity Recognition (WHAR) is a prominent research area within ubiquitous computing. Multi-sensor synchronous measurement has proven to be more effective for WHAR than using a single sensor. However, existing WHAR methods use shared convolutional kernels for indiscriminate temporal feature extraction across each sensor variable, which fails to effectively capture spatio-temporal relationships of intra-sensor and inter-sensor variables. We propose the DecomposeWHAR model consisting of a decomposition phase and a fusion phase to better model the relationships between modality variables. The decomposition creates high-dimensional representations of each intra-sensor variable through the improved Depth Separable Convolution to capture local temporal features while preserving their unique characteristics. The fusion phase begins by capturing relationships between intra-sensor variables and fusing their features at both the channel and variable levels. Long-range temporal dependencies are modeled using the State Space Model (SSM), and later cross-sensor interactions are dynamically captured through a self-attention mechanism, highlighting inter-sensor spatial correlations. Our model demonstrates superior performance on three widely used WHAR datasets, significantly outperforming state-of-the-art models while maintaining acceptable computational efficiency.

Traffic prediction provides vital support for urban traffic management and has received extensive research interest. By virtue of the ability to effectively learn spatial and temporal dependencies from a global view, Transformers have achieved superior performance in long-term traffic prediction. However, existing methods usually underrate the complex spatio-temporal entanglement in long-range sequences. Compared with purely temporal entanglement, spatio-temporal data emphasizes the entangled dynamics under the restrictions of traffic networks, which brings additional difficulties. Moreover, the computational costs of spatio-temporal Transformers scale quadratically as the sequence length grows, limiting their applications on long-range and large-scale scenarios. To address these problems, we propose a decomposed spatio-temporal Mamba (DST-Mamba) for traffic prediction. We aim to apply temporal decomposition to the entangled sequences and obtain the seasonal and trend parts. Shifting from the temporal view to the spatial view, we leverage Mamba, a state space model with near-linear complexity, to capture seasonal variations in a node-centric manner. Meanwhile, multi-scale trend information is extracted and aggregated by simple linear layers. Such combination equips DST-Mamba with superior capability to model long-range spatio-temporal dependencies while remaining efficient compared with Transformers. Experimental results across five real-world datasets demonstrate that DST-Mamba can capture both local fluctuations and global trends within traffic patterns, achieving state-of-the-art performance with favorable efficiency.

Algorithms and hardware for solving quadratic unconstrained binary optimization (QUBO) problems have made significant recent progress. This advancement has focused attention on formulating combinatorial optimization problems as quadratic polynomials. To improve the performance of solving large QUBO problems, it is essential to minimize the number of binary variables used in the objective function. In this paper, we propose a QUBO formulation that offers a bit capacity advantage over conventional quadratization techniques. As a key application, this formulation significantly reduces the number of binary variables required for score-based Bayesian network structure learning. Experimental results on 16 instances, ranging from 37 to 223 variables, demonstrate that our approach requires fewer binary variables than quadratization by orders of magnitude. Moreover, an annealing machine that implement our formulation have outperformed existing algorithms in score maximization.

We consider the problem of learning stable matchings with unknown preferences in a decentralized and uncoordinated manner, where ``decentralized" means that players make decisions individually without the influence of a central platform, and ``uncoordinated" means that players do not need to synchronize their decisions using pre-specified rules. First, we provide a game formulation for this problem with known preferences, where the set of pure Nash equilibria (NE) coincides with the set of stable matchings, and mixed NE can be rounded to a stable matching. Then, we show that for hierarchical markets, applying the exponential weight (EXP) learning algorithm to the stable matching game achieves logarithmic regret in a fully decentralized and uncoordinated fashion. Moreover, we show that EXP converges locally and exponentially fast to a stable matching in general matching markets. We complement our results by introducing another decentralized and uncoordinated learning algorithm that globally converges to a stable matching with arbitrarily high probability.

This paper studies decentralized optimization over a compact submanifold within a communication network of n nodes, where each node possesses a smooth non-convex local cost function, and the goal is to jointly minimize the sum of these local costs. We focus particularly on the online setting, where local data is processed in real-time as it streams in, without the need for full data storage. We propose a decentralized projected Riemannian stochastic recursive momentum (DPRSRM) method that employs local hybrid stochastic gradient estimators and uses the network to track the global gradient. DPRSRM achieves an oracle complexity of O(epsilon^(-3/2)), outperforming existing methods that have at most O(epsilon^(-2)) complexity. Our method requires only O(1) gradient evaluations per iteration for each local node and does not require restarting with a large batch gradient. Furthermore, we demonstrate the effectiveness of our proposed methods compared to state-of-the-art ones through numerical experiments on principal component analysis problems and low-rank matrix completion.

Federated Learning (FL) trains a shared model using data and computation power on distributed agents coordinated by a central server. Decentralized FL (DFL) utilizes local model exchange and aggregation between agents to reduce the communication and computation overheads on the central server. However, when agents are mobile, the communication opportunity between agents can be sporadic, largely hindering the convergence and accuracy of DFL. In this paper, we propose Cached Decentralized Federated Learning (Cached-DFL) to investigate delay-tolerant model spreading and aggregation enabled by model caching on mobile agents. Each agent stores not only its own model, but also models of agents encountered in the recent past. When two agents meet, they exchange their own models as well as the cached models. Local model aggregation utilizes all models stored in the cache. We theoretically analyze the convergence of Cached-DFL, explicitly taking into account the model staleness introduced by caching. We design and compare different model caching algorithms for different DFL and mobility scenarios. We conduct detailed case studies in a vehicular network to systematically investigate the interplay between agent mobility, cache staleness, and model convergence. In our experiments, Cached-DFL converges quickly, and significantly outperforms DFL without caching.

We propose a decentralized market model in which agents can negotiate bilateral contracts. This builds on a similar, but centralized, model of trading networks introduced by Hatfield et al. in 2013. Prior work has established that fully-substitutable preferences guarantee the existence of competitive equilibria which can be centrally computed. Our motivation comes from the fact that prices in markets such as over-the-counter markets and used car markets arise from decentralized negotiation among agents, which has left open an important question as to whether equilibrium prices can emerge from agent-to-agent bilateral negotiations. We design a best response dynamic intended to capture such negotiations between market participants. We assume fully substitutable preferences for market participants. In this setting, we provide proofs of convergence for sparse markets (covering many real world markets of interest), and experimental results for more general cases, demonstrating that prices indeed reach equilibrium, quickly, via bilateral negotiations. Our best response dynamic, and its convergence behavior, forms an important first step in understanding how decentralized markets reach, and retain, equilibrium.

Human multimodal language understanding (MLU) is an indispensable component of expression analysis (e.g., sentiment or humor) from heterogeneous modalities, including visual postures, linguistic contents, and acoustic behaviours. Existing works invariably focus on designing sophisticated structures or fusion strategies to achieve impressive improvements. Unfortunately, they all suffer from the subject variation problem due to data distribution discrepancies among subjects. Concretely, MLU models are easily misled by distinct subjects with different expression customs and characteristics in the training data to learn subject-specific spurious correlations, limiting performance and generalizability across new subjects. Motivated by this observation, we introduce a recapitulative causal graph to formulate the MLU procedure and analyze the confounding effect of subjects. Then, we propose SuCI, a simple yet effective causal intervention module to disentangle the impact of subjects acting as unobserved confounders and achieve model training via true causal effects. As a plug-and-play component, SuCI can be widely applied to most methods that seek unbiased predictions. Comprehensive experiments on several MLU benchmarks clearly show the effectiveness of the proposed module.

Knowledge distillation transfers "dark knowledge" from a large teacher model to a smaller student model, yielding a highly efficient network. To improve network's generalization ability, existing works use a larger temperature coefficient for knowledge distillation. Nevertheless, these methods may lower the target category's confidence and lead to ambiguous recognition of similar samples. To mitigate this issue, some studies introduce intra-batch distillation to reduce prediction discrepancy. However, these methods overlook the inconsistency between background information and the target category, which may increase prediction bias due to noise disturbance. Additionally, label imbalance from random sampling and batch size can undermine network generalization reliability. To tackle these challenges, we propose a simple yet effective Intra-class Knowledge Distillation (IKD) method that facilitates knowledge sharing within the same class to ensure consistent predictions. First, we initialize the matrix and the vector to store logits and class counts provided by the teacher, respectively. Then, in the first epoch, we calculate the sum of logits and sample counts per class and perform KD to prevent knowledge omission. Finally, in subsequent training, we update the matrix to obtain the average logits and compute the KL divergence between the student's output and the updated matrix according to the label index. This process ensures intra-class consistency and improves the student's performance. Furthermore, this method theoretically reduces prediction bias by ensuring intra-class consistency. Extensive experiments on the CIFAR-100, ImageNet-1K, and Tiny-ImageNet datasets validate the superiority of IKD.

All-in-one image restoration is a fundamental low-level vision task with significant real-world applications. The primary challenge lies in addressing diverse degradations within a single model. While current methods primarily exploit task prior information to guide the restoration models, they typically employ uniform multi-task learning, overlooking the heterogeneity in model optimization across different degradation tasks. To eliminate the bias, we propose a task-aware optimization strategy, that introduces adaptive task-specific regularization for multi-task image restoration learning. Specifically, our method dynamically weights and balances losses for different restoration tasks during training, encouraging the implementation of the most reasonable optimization route. In this way, we can achieve more robust and effective model training. Notably, our approach can serve as a plug-and-play strategy to enhance existing models without requiring modifications during inference. Extensive experiments in diverse all-in-one restoration settings demonstrate the superiority and generalization of our approach. For example, AirNet retrained with TUR achieves average improvements of 1.16 dB on three distinct tasks and 1.81 dB on five distinct all-in-one tasks. These results underscore TUR's effectiveness in advancing the SOTAs in all-in-one image restoration, paving the way for more robust and versatile image restoration.

The strategy of selecting ``most informative'' hard samples in active learning has proven a boon for alleviating the challenges of few-shot learning and costly data annotation in deep learning. However, this very preference towards hard samples engenders bias issues, thereby impeding the full potential of active learning. It has witnessed an increasing trend to mitigate this stubborn problem, yet most neglect the quantification of bias itself and the direct rectification of dynamically evolving biases. Revisiting the bias issue, this paper presents an active learning approach based on the Variational Gradient Rectifier (VaGeRy). First, we employ variational methods to quantify bias at the level of latent state representations. Then, harnessing historical training dynamics, we introduce Uncertainty Consistency Regularization and Fluctuation Restriction, which asynchronously iterate to rectify gradient backpropagation. Extensive experiments demonstrate that our proposed methodology effectively counteracts bias phenomena in a majority of active learning scenarios

Large Language Models (LLMs) may suffer from hallucinations in real-world applications due to the lack of relevant knowledge. In contrast, knowledge graphs encompass extensive, multi-relational structures that store a vast array of symbolic facts. Consequently, integrating LLMs with knowledge graphs has been extensively explored, with Knowledge Graph Question Answering (KGQA) serving as a critical touchstone for the integration. This task requires LLMs to answer natural language questions by retrieving relevant triples from knowledge graphs. However, existing methods face two significant challenges: *excessively long reasoning paths distracting from the answer generation*, and *false-positive relations hindering the path refinement*. In this paper, we propose an iterative interactive KGQA framework that leverages the interactive learning capabilities of LLMs to perform reasoning and Debating over Graphs (DoG). Specifically, DoG employs a subgraph-focusing mechanism, allowing LLMs to perform answer trying after each reasoning step, thereby mitigating the impact of lengthy reasoning paths. On the other hand, DoG utilizes a multi-role debate team to gradually simplify complex questions, reducing the influence of false-positive relations. This debate mechanism ensures the reliability of the reasoning process. Experimental results on five public datasets demonstrate the effectiveness and superiority of our architecture. Notably, DoG outperforms the state-of-the-art method ToG by 23.7% and 9.1% in accuracy on WebQuestions and GrailQA, respectively. Furthermore, the integration experiments with various LLMs on the mentioned datasets highlight the flexibility of DoG.

Common methods for aligning already-capable models with desired behavior rely on the ability of humans to provide supervision. However, future superhuman models will surpass the capability of humans. Therefore, humans will only be able to weakly supervise superhuman models. This expected deficiency of human evaluation would weaken the safety of future AI systems. Scalable oversight and weak-to-strong generalization are two complementary approaches to tackle this issue. In this paper, we attempt to combine the strengths of these two approaches to further improve alignment. Specifically, we investigate ways of improving human supervision with a strong pretrained model and then supervise the strong model with enhanced weak human supervision. To make iterative empirical progress, we consider an analogy: can we use a strong model to improve weak model supervision and then use it to supervise the strong model? We empirically test it by finetuning a small weak model on ground truth labels with the additional help from a large strong model, and then finetuning the strong model on labels generated by the weak model. We find that debate can assist a weak model in extracting trustworthy information from an untrustworthy strong model, which provides leverage as context on samples when training a weak model. We also show that an ensemble of weak models helps exploit long arguments generated by strong model debaters and obtain a more robust supervision estimate. Extensive experiments on the OpenAI weak-to-strong NLP benchmarks show that the combination approach leads to better alignment, which indicates that debate has the potential to help weak-to-strong generalization.

Exploring the correlations between medical features is essential for extracting patient health patterns from electronic health records (EHR) data, and strengthening medical predictions and decision-making. To constrain the hypothesis space of pure data-driven deep learning in the context of limited annotated data, a common trend is to incorporate external knowledge, especially knowledge priors related to personalized health contexts, to optimize model training. However, most existing methods lack flexibility and are constrained by the uncertainties brought about by fixed feature correlation priors. In addition, in utilizing knowledge, these methods overlook the knowledge informative for personalized healthcare. To this end, we propose DearLLM, a novel and effective framework that leverages feature correlations deduced by large language models (LLMs) to enhance personalized healthcare. Concretely, DearLLM captures and learns quantitative correlations between medical features by calculating the conditional perplexity of LLMs’ deduction based on personalized patient backgrounds. Then, DearLLM enhances healthcare predictions by emphasizing knowledge that carries unique patient information through a feature-frequency-aware graph pooling method. Extensive experiments on two real-world benchmark datasets show significant performance gains brought by DearLLM. Furthermore, the discovered findings align well with medical literature, offering meaningful clinical interpretations.

Reconstruction under adverse rainy conditions poses significant challenges due to reduced visibility and the distortion of visual perception. These conditions can severely impair the quality of geometric maps, which is essential for applications ranging from autonomous planning to environmental monitoring. In response to these challenges, this study introduces the novel task of 3D Reconstruction in Rainy Environments (3DRRE), specifically designed to address the complexities of reconstructing 3D scenes under rainy conditions. To benchmark this task, we construct the HydroViews dataset that comprises a diverse collection of both synthesized and real-world scene images characterized by various intensities of rain streaks and raindrops. Furthermore, we propose DeRainGS, the first 3DGS method tailored for reconstruction in adverse rainy environments. Extensive experiments across a wide range of rain scenarios demonstrate that our method delivers state-of-the-art performance, remarkably outperforming existing occlusion-free methods by a large margin.

Recent years have witnessed the rise of Neural-enhanced Video Streaming (NeVS), which integrates neural restoration models into video codecs for higher compression-restoration performance. Despite its benefit, existing work has not well explored the full potential of NeVS paradigm, due to: (1) post-streaming restoration by decoder while lacking the proactive collaboration of encoder, (2) end-to-end optimization based on conventional rate-distortion theory, which has been verified that low distortion is not always a synonym for high perceptual quality, and (3) coupled design for domain-specific tasks that cannot generalize to various video codecs. Observing these limitations, our objective is not to incrementally present an improved restoration model. Instead, we focus on the encoder-decoder synergy, i.e., the codec, which is non-trivial since it inherently strikes the rate-distortion-perception trade-off of NeVS. Aiming at this target, we propose the Diffusion-enhanced Neural Codec (DeNC), a plug-and-play module for current NeVS paradigm, to significantly reduce the required bitrates while preserving high perceptual quality of restored videos. Our key design is twofold. First, DeNC improves the encoder's compression efficiency by simultaneously reducing the resolution and color bit-depth of frame referencing. Second, DeNC empowers the decoder with perception-oriented restoration capability by making its diffusion-based restoration process aware of the encoder's compression conditions. Real-world evaluations show that DeNC improves compression ratios with nearly an order of magnitude and achieves much higher restoration quality (e.g., 93+ VMAF and 23% higher MOS) over the latest baselines.

As a long-range prior, motion consensus essentially forces the overall spatial transformation between a pair of images to be smooth and consistent, which is naturally well-suited for two-view correspondence learning. However, such precious property remains under-explored by most existing studies due to the modeling challenges posed by the sparsity and uneven distributions of putative correspondences. In this paper, we propose DeMo, a novel and cutting-edge network for outlier rejection, which possesses the capacity to fully capture global motion consensus clues by way of consensus interpolation over the entire high-dimensional motion field generated by putative correspondences. Specifically, through incorporating regularization techniques into a Reproducing Kernel Hilbert Space (RKHS), a concise interpolation formula can be derived for the high-dimensional motion field, which inherently allows a closed-form solution. Subsequently, learnable deep kernels are collaboratively used to flexibly and efficiently capture the relationships between global inputs, thus maintaining the entire motion field consensus. In addition, to remedy the cubic computational overhead of explicit interpolation, a scene-adaptive sampling strategy is introduced, which implicitly selects the more scene-representative motions, reducing the computational complexity of motion consensus interpolation to be approximately linear while maintaining the accuracy. Moreover, to deal with underlying depth discontinuities caused by complicated scene variations, a local consensus complementation block is designed, which maintains local bilateral consensus across both feature and spatial channels. Without bells and whistles, DeMo achieves superior performance in various geometric tasks, including relative pose estimation, homography estimation, and visual localization.