AAAI 2025 Papers

Reinforcement Learning (RL) has emerged as a powerful paradigm for sequential decision-making with numerous real-world applications. However, in practical environments such as recommender systems, search engines, and LLMs, RL algorithms must efficiently learn from biased human feedback that may be subject to corruption. In this talk, I will present our recent efforts in developing robust RL algorithms that can provably effectively handle such challenging scenarios. First, I will introduce our works on reinforcement learning from biased click feedback in ranking. While previous approaches typically relied on strong assumptions about human click behavior (formalized as click models) and required specialized debiasing methods for different models, we propose a novel unified framework that formulates the ranking process under general click models as a Markov Decision Process, enabling the development of a click model-agnostic RL algorithm. Second, I will introduce the fundamental vulnerability of bandits and reinforcement learning under corrupted feedback. Our theoretical analysis provides complete necessity and sufficiency characterizations of the attackability of linear bandits and linear RL, revealing their intrinsic robustness and limitations. Lastly, I will discuss our recent works on improving RL finetuning for LLMs, including sample efficient off-policy RLHF and solving the gradient entanglement issue in margin-based alignment methods.

Machine unlearning is becoming increasingly important as deep models become more prevalent, particularly when there are frequent requests to remove the influence of specific training data due to privacy concerns or erroneous sensing signals. Spatial-temporal Graph Neural Networks, in particular, have been widely adopted in real-world applications that demand efficient unlearning, yet research in this area remains in its early stages. In this paper, we introduce STEPS, a framework specifically designed to address the challenges of spatio-temporal graph unlearning. Our results demonstrate that STEPS not only ensures data continuity and integrity but also significantly reduces the time required for unlearning, while minimizing the accuracy loss in the new model compared to a model with 0% unlearning.

Fractional-order differential equations (FDEs) enhance traditional differential equations by extending the order of differential operators from integers to real numbers, offering greater flexibility in modeling complex dynamic systems with nonlocal characteristics. Recent progress at the intersection of FDEs and deep learning has catalyzed a new wave of innovative models, demonstrating the potential to address challenges such as graph representation learning. However, training neural FDEs has primarily relied on direct differentiation through forward-pass operations in FDE numerical solvers, leading to increased memory usage and computational complexity, particularly in large-scale applications. To address these challenges, we propose a scalable adjoint backpropagation method for training neural FDEs by solving an augmented FDE backward in time, which substantially reduces memory requirements. This approach provides a practical neural FDE toolbox and holds considerable promise for diverse applications. We demonstrate the effectiveness of our method in several tasks, achieving performance comparable to baseline models while significantly reducing computational overhead.

Graph neural networks (GNNs) have gained considerable attention in recent years for traffic flow prediction due to their ability to learn spatio-temporal pattern representations through a graph-based message-passing framework. Although GNNs have shown great promise in handling traffic datasets, their deployment in real-life applications has been hindered by scalability constraints arising from high-order message passing. Additionally, the over-smoothing problem of GNNs may lead to indistinguishable region representations as the number of layers increases, resulting in performance degradation. To address these challenges, we propose a new knowledge distillation paradigm termed LightST that transfers spatial and temporal knowledge from a high-capacity teacher to a lightweight student. Specifically, we introduce a spatio-temporal knowledge distillation framework that helps student MLPs capture graph-structured global spatio-temporal patterns while alleviating the over-smoothing effect with adaptive knowledge distillation. Extensive experiments verify that LightST significantly speeds up traffic flow predictions by 5X to 40X compared to state-of-the-art spatio-temporal GNNs, all while maintaining superior accuracy.

Self-supervised video hashing (SSVH) is a practical task in video indexing and retrieval. Although Transformers are predominant in SSVH for their impressive temporal modeling capabilities, they often suffer from computational and memory inefficiencies. Drawing inspiration from Mamba, an advanced state-space model, we explore its potential in SSVH to achieve a better balance between efficacy and efficiency. We introduce S5VH, a Mamba-based video hashing model with an improved self-supervised learning paradigm. Specifically, we design bidirectional Mamba layers for both the encoder and decoder, which are effective and efficient in capturing temporal relationships thanks to the data-dependent selective scanning mechanism with linear complexity. In our learning strategy, we transform global semantics in the feature space into semantically consistent and discriminative hash centers, followed by a center alignment loss as a global learning signal. Our self-local-global (SLG) paradigm significantly improves learning efficiency, leading to faster and better convergence. Extensive experiments demonstrate S5VH's improvements over state-of-the-art methods, superior transferability, and scalable advantages in inference efficiency.

The study of enhancing model robustness against adversarial examples has become increasingly critical in the security of deep learning, leading to the development of numerous adversarial defense techniques. While these defense methods have shown promise in mitigating the impact of adversarial perturbations, evaluating their effectiveness remains a critical challenge. The recently introduced AutoAttack technique has been recognized as a standardized method for assessing model robustness. However, the computational demands of the AutoAttack method significantly limits its applicability, underscoring the urgent need for efficient evaluation techniques. To address this challenge, we propose a novel and efficient evaluation framework based on strategic constraint relaxation. Our key insight is that temporarily expanding the adversarial perturbation bounds during the attack process can help discover more effective adversarial examples. Based on this insight, we develop the Constraint Relaxation Attack (CR Attack) method, which systematically relaxes and resets perturbation constraints during optimization. Extensive experiments on 105 robust models show that CR Attack outperforms AutoAttack in both attack success rate and efficiency, reducing forward and backward propagation time by 38.3× and 15.9× respectively. Through comprehensive analysis, we validate that the constraint relaxation mechanism is crucial for the method's effectiveness.

In many human-robot collaboration and multi-agent tasks, it is vital to model the partners and estimate their objectives to efficiently collaborate/interact with them. While learning from demonstrations is the most common approach for this, it is very data-hungry, which we cannot afford in many settings including robotics, and demonstrations are unreliable in a surprisingly large number of domains, including those we think humans perform reasonably well, e.g., driving. In this talk, I will start with introducing comparison-based feedback and explain why it does not suffer from most of the problems that demonstrations have, but is still data-hungry. To address this problem, I will propose comparative language based feedback and active learning techniques, which will result in (1) a new type of human feedback, and (2) an active querying algorithm that optimizes the information the AI agent will elicit from the human. I will conclude the talk by discussing what other types of human feedback exist, e.g., interventions or hand gestures, and how we can incorporate them into the existing learning algorithms.

In this paper, we study reinforcement learning in Markov Decision Processes with Probabilistic Reward Machines (PRMs), a form of non-Markovian reward commonly found in robotics tasks. We design an algorithm for PRMs that achieves a regret bound of Õ((HOAT)^(1/2) + H²O²A^(3/2) + H(T)^(1/2)), where H is the time horizon, O is the number of observations, A is the number of actions, and T is the number of time steps. This result improves over the best-known bound, Õ(H(OAT)^(1/2)), for MDPs with Deterministic Reward Machines (DRMs), a special case of PRMs. When T ≥ H³O³A² and OA ≥ H, our regret bound leads to a regret of Õ((HOAT)^(1/2)), which matches the established lower bound of Ω((HOAT)^(1/2)) for MDPs with DRMs up to a logarithmic factor. To the best of our knowledge, this is the first efficient algorithm for PRMs. Additionally, we present a new simulation lemma for non-Markovian rewards, which enables reward-free exploration for any non-Markovian reward given access to an approximate planner. Complementing our theoretical findings, we show through extensive experimental evaluations that our algorithm indeed outperforms prior methods in various PRM environments.

While Reinforcement Learning (RL) agents can successfully learn to handle complex tasks, effectively generalizing acquired skills to unfamiliar settings remains a challenge. One of the reasons behind this is the visual encoder used are task-dependent, preventing effective feature extraction in different settings. To address this issue, recent studies have tried to pretrain encoders with diverse visual inputs in order to improve their performance. However, they rely on existing pretrained encoders without further exploring the impact of pretraining period. In this work, we propose APE: efficient reinforcement learning through Adaptively Pretrained visual Encoder—a framework that utilizes adaptive augmentation strategy during the pretraining phase and extracts useful features with only a few interactions within the task environments in the policy learning period. Experiments are conducted across various domains, including DeepMind Control Suite, Atari Games and Memory Maze benchmarks, to verify the effectiveness of our method. Results show that mainstream RL methods, such as DreamerV3 and DrQ-v2, achieve state-of-the-art performance when equipped with APE. In addition, APE significantly improves the sampling efficiency during learning, approaching the efficiency of state-based method using only visual inputs in several control tasks. These findings demonstrate the potential of adaptive pretraining of encoder in enhancing the generalization ability and efficiency of visual RL algorithms.

Neuro-Symbolic (NeSy) AI could be regarded as an analogy to human dual-process cognition, modeling the intuitive System 1 with neural networks and the algorithmic System 2 with symbolic reasoning. However, for complex learning targets, NeSy systems often generate outputs inconsistent with domain knowledge and it is challenging to rectify them. Inspired by the human Cognitive Reflection, which promptly detects errors in our intuitive response and revises them by invoking the System 2 reasoning, we propose to improve NeSy systems by introducing Abductive Reflection (ABL-Refl) based on the Abductive Learning (ABL) framework. ABL-Refl leverages domain knowledge to abduce a reflection vector during training, which can then flag potential errors in the neural network outputs and invoke abduction to rectify them and generate consistent outputs during inference. ABL-Refl is highly efficient in contrast to previous ABL implementations. Experiments show that ABL-Refl outperforms state-of-the-art NeSy methods, achieving excellent accuracy with fewer training resources and enhanced efficiency.

In time-lapse microscopy, inherent noise significantly limits imaging sensitivity and increases measurement uncertainty. Due to the scarcity of clean data, zero-shot approaches have emerged as highly data-efficient solutions for microscopy denoising. However, existing methods typically process video frames independently, resulting in long training times and issues such as temporal noise and over-smoothing. In this paper, we introduce MDSR-Zero, a zero-shot online learning method designed for plug-and-play noise suppression and super-resolution of microscopy videos. Our approach leverages an efficient online training strategy that reuses denoising models from previous frames. By treating the video as a continuous stream, our model significantly reduces training time and ensures temporally consistent denoising. Additionally, we propose a novel loss function tailored for denoising in the context of super-resolution, which enhances the detail in the denoised results. Extensive experiments on both synthetic and real-world noise demonstrate that our method achieves state-of-the-art performance among zero-shot denoising approaches and is competitive with self-supervised methods. Notably, our method can reduce training time by up to 10x compared to the previous SOTA method.

3D color lookup tables (LUTs) enable precise color manipulation by mapping input RGB values to specific output RGB values. 3D LUTs are instrumental in various applications, including video editing, in-camera processing, photographic filters, computer graphics, and color processing for displays. While an individual LUT does not incur a high memory overhead, software and devices may need to store dozens to hundreds of LUTs that can take over 100 MB. This work aims to develop a neural network architecture that can encode hundreds of LUTs in a single compact representation. To this end, we propose a model with a memory footprint of less than 0.25 MB that can reconstruct 512 LUTs with only minor color distortion (ΔE ≤ 2.0 on average) over the entire color gamut. We also show that our network can weight colors to provide further quality gains on natural image colors (ΔE ≤ 1.0 on average). Finally, we show that minor modifications to the network architecture enable a bijective encoding that produces LUTs that are invertible, allowing for reverse color processing.

To unbiasedly evaluate multiple target policies, the dominant approach among RL practitioners is to run and evaluate each target policy separately. However, this evaluation method is far from efficient because samples are not shared across policies, and running target policies to evaluate themselves is actually not optimal. In this paper, we address these two weaknesses by designing a tailored behavior policy to reduce the variance of estimators across all target policies. Theoretically, we prove that executing this behavior policy with manyfold fewer samples outperforms on-policy evaluation on every target policy under characterized conditions. Empirically, we show our estimator has a substantially lower variance compared with previous best methods and achieves state-of-the-art performance in a broad range of environments.

The ability to autonomously explore and resolve tasks with minimal human guidance is crucial for the self-development of embodied intelligence. Although reinforcement learning methods can largely ease human effort, it's challenging to design reward functions for real-world tasks, especially for high-dimensional robotic control, due to complex relationships among joints and tasks. Recent advancements large language models (LLMs) enable automatic reward function design. However, approaches evaluate reward functions by re-training policies from scratch placing an undue burden on the reward function, expecting it to be effective throughout the whole policy improvement process. We argue for a more practical strategy in robotic autonomy, focusing on refining existing policies with policy-dependent reward functions rather than a universal one. To this end, we propose a novel reward-policy co-evolution framework where the reward function and the learned policy benefit from each other's progressive on-the-fly improvements, resulting in more efficient and higher-performing skill acquisition. Specifically, the reward evolution process translates the robot's previous best reward function, descriptions of tasks and environment into text inputs. These inputs are used to query LLMs to generate a dynamic amount of reward function candidates, ensuring continuous improvement at each round of evolution. For policy evolution, our method generates new policy populations by hybridizing historically optimal and random policies. Through an improved Bayesian optimization, our approach efficiently and robustly identifies the most capable and plastic reward-policy combination, which then proceeds to the next round of co-evolution. Despite using less data, our approach demonstrates an average normalized improvement of 95.3\% across various high-dimensional robotic skill learning tasks.

Most indoor depth completion tasks rely on convolutional auto-encoders to reconstruct depth images, especially in areas with significant missing values. While traditional convolution treats valid and missing pixels equally, Partial Convolution (PConv) has mitigated this limitation. However, PConv fails to distinguish the varying degree of invalidity across different missing areas, which highlights the need for a more refined strategy. To solve this problem, we propose a novel system for indoor depth completion tasks that leverages Mask-adaptive Gated Convolution (MagaConv). MagaConv utilizes gated signals to selectively apply convolution kernels based on the characteristics of missing depth data. These gating signals are generated using shared convolution kernels that jointly process depth features and corresponding masks, ensuring coherent weight optimization. Additionally, the mask undergoes iterative updates according to predefined rules. To improve the fusion of depth and color information, we introduce a Bi-directional Aligning Projection (Bid-AP) module, which utilizes a bi-directional projection scheme with global spatial-channel attention mechanisms to filter out depth-irrelevant features from other modalities. Extensive experiments on popular benchmarks, including NYU-Depth V2, DIML, and SUN RGB-D, demonstrate that our model outperforms state-of-the-art methods in both accuracy and efficiency.

Diffusion models (DMs) have demonstrated great potential in the field of adversarial robustness, where DM-based defense methods can achieve superior defense capability without adversarial training. However, they all require huge computational costs due to the usage of large-scale pre-trained DMs, making it difficult to conduct full evaluation under strong attacks and compare with traditional CNN-based methods. Simply reducing the network size and timesteps in DMs could significantly harm the image generation quality, which invalidates previous frameworks. To alleviate this issue, we redesign the diffusion framework from generating high-quality images to predicting distinguishable image labels. Specifically, we employ an image translation framework to learn many-to-one mapping from input samples to designed orthogonal image labels. Based on this framework, we introduce an efficient Image-to-Image diffusion classifier with a pruned U-Net structure and reduced diffusion timesteps. Besides the framework, we redesign the optimization objective of DMs to fit the target of image classification, where a new classification loss is incorporated in the DM-based image translation framework to distinguish the generated label from those of other classes. We conduct sufficient evaluations of the proposed classifier under various attacks on popular benchmarks. Extensive experiments show that our method achieves better adversarial robustness with fewer computational costs than DM-based and CNN-based methods.

In this paper, we present a new image similarity search algorithm designed to enhance traditional information retrieval(IR) by adding an image search capability. Our approach uses a quadtree data structure to organize image data, significantly reducing search space and improving retrieval efficiency. We describe an indexing strategy and two query algorithms that can be implemented in any IR system. We tested our method on a 70K material microscopy image dataset, achieving a 25 times improvement in retrieval speed with only a 20% reduction in ranking accuracy.

This paper presents a novel framework for multi-armed bandit problems with side-observations and switching constraints, which arises in a range of real-world applications such as robotic. To address the challenges of effectively utilizing graph-structured observations while adhering to graph constraints, we design graph-agnostic and graph-aware algorithms tailored to this new setting. Specifically, our graph-agnostic algorithm selects nodes with the highest upper confidence bound without prior knowledge of feedback probabilities, while minimizing switching costs using offline shortest path planning and the doubling trick. If the graph structure and associated probability matrix are known, our graph-aware algorithm plans the exploration step using a linear programming approach and eliminates suboptimal nodes iteratively. We rigorously analyze the performance of our proposed algorithms, providing near-optimal minimax and instance-dependent regret upper bounds. Our analysis shows that our algorithms outperform generic reinforcement learning methods in terms of both regret and computational efficiency. Extensive numerical experiments on various types of graphs, including two real-world datasets, demonstrate the efficacy of our proposed methods and their advantages over benchmark methods in graph bandit settings.

Rendering dynamic scenes from monocular videos is a crucial yet challenging task. The recent deformable Gaussian Splatting has emerged as a robust solution to represent real-world dynamic scenes. However, it often leads to heavily redundant Gaussians, attempting to fit every training view at various time steps, leading to slower rendering speeds. Additionally, the attributes of Gaussians in static areas are time-invariant, making it unnecessary to model every Gaussian, which can cause jittering in static regions. In practice, the primary bottleneck in rendering speed for dynamic scenes is the number of Gaussians. In response, we introduce Efficient Dynamic Gaussian Splatting (EDGS), which represents dynamic scenes via sparse time-variant attribute modeling. Our approach formulates dynamic scenes using a sparse anchor-grid representation, with the motion flow of dense Gaussians calculated via a classical kernel representation. Furthermore, we propose an unsupervised strategy to efficiently filter out anchors corresponding to static areas. Only anchors associated with deformable objects are input into MLPs to query time-variant attributes. Experiments on two real-world datasets demonstrate that our EDGS significantly improves the rendering speed with superior rendering quality compared to previous state-of-the-art methods.