AAAI 2024 论文列表

Generating stylized talking head with diverse head motions is crucial for achieving natural-looking videos but still remains challenging. Previous works either adopt a regressive method to capture the speaking style, resulting in a coarse style that is averaged across all training data, or employ a universal network to synthesize videos with different styles which causes suboptimal performance. To address these, we propose a novel dynamic-weight method, namely Say Anything with Any Style (SAAS), which queries the discrete style representation via a generative model with a learned style codebook. Specifically, we develop a multi-task VQ-VAE that incorporates three closely related tasks to learn a style codebook as a prior for style extraction. This discrete prior, along with the generative model, enhances the precision and robustness when extracting the speaking styles of the given style clips. By utilizing the extracted style, a residual architecture comprising a canonical branch and style-specific branch is employed to predict the mouth shapes conditioned on any driving audio while transferring the speaking style from the source to any desired one. To adapt to different speaking styles, we steer clear of employing a universal network by exploring an elaborate HyperStyle to produce the style-specific weights offset for the style branch. Furthermore, we construct a pose generator and a pose codebook to store the quantized pose representation, allowing us to sample diverse head motions aligned with the audio and the extracted style. Experiments demonstrate that our approach surpasses state-of-the-art methods in terms of both lip-synchronization and stylized expression. Besides, we extend our SAAS to video-driven style editing field and achieve satisfactory performance as well.

In this paper, we present SasWOT, the first training-free Semantic segmentation Architecture Search (SAS) framework via an auto-discovery proxy. Semantic segmentation is widely used in many real-time applications. For fast inference and memory efficiency, Previous SAS seeks the optimal segmenter by differentiable or RL Search. However, the significant computational costs of these training-based SAS limit their practical usage. To improve the search efficiency, we explore the training-free route but empirically observe that the existing zero-cost proxies designed on the classification task are sub-optimal on the segmentation benchmark. To address this challenge, we develop a customized proxy search framework for SAS tasks to augment its predictive capabilities. Specifically, we design the proxy search space based on the some observations: (1) different inputs of segmenter statistics can be well combined; (2) some basic operators can effectively improve the correlation. Thus, we build computational graphs with multiple statistics as inputs and different advanced basis arithmetic as the primary operations to represent candidate proxies. Then, we employ an evolutionary algorithm to crossover and mutate the superior candidates in the population based on correlation evaluation. Finally, based on the searched proxy, we perform the segmenter search without candidate training. In this way, SasWOT not only enables automated proxy optimization for SAS tasks but also achieves significant search acceleration before the retrain stage. Extensive experiments on Cityscapes and CamVid datasets demonstrate that SasWOT achieves superior trade-off between accuracy and speed over several state-of-the-art techniques. More remarkably, on Cityscapes dataset, SasWOT achieves the performance of 71.3% mIoU with the speed of 162 FPS.

Multi-Object Tracking (MOT) is a cornerstone operator for video surveillance applications. To enable real-time processing of large-scale live video streams, we study an interesting scenario called down-sampled MOT, which performs object tracking only on a small subset of video frames. The problem is challenging for state-of-the-art MOT methods, which exhibit significant performance degradation under high frame reduction ratios. In this paper, we devise a sampling-resilient tracker with a novel sparse-observation Kalman filter (SOKF). It integrates an LSTM network to capture non-linear and dynamic motion patterns caused by sparse observations. Since the LSTM-based state transition is not compatible with the original noise estimation mechanism, we propose new estimation strategies based on Bayesian neural networks and derive the optimal Kalman gain for SOKF. To associate the detected bounding boxes robustly, we also propose a comprehensive similarity metric that systematically integrates multiple spatial matching signals. Experiments on three benchmark datasets show that our proposed tracker achieves the best trade-off between efficiency and accuracy. With the same tracking accuracy, we reduce the total processing time of ByteTrack by 2× in MOT17 and 3× in DanceTrack.

The feedback arc set problem is one of the most fundamental and well-studied ranking problems where n objects are to be ordered based on their pairwise comparison. The problem enjoys several efficient approximation algorithms in the offline setting. Unfortunately, online there are strong lower bounds on the competitive ratio establishing that no algorithm can perform well in the worst case. This paper introduces a new beyond-worst-case model for online feedback arc set. In the model, a sample of the input is given to the algorithm offline before the remaining instance is revealed online. This models the case in practice where yesterday's data is available and is similar to today's online instance. This sample is drawn from a known distribution which may not be uniform. We design an online algorithm with strong theoretical guarantees. The algorithm has a small constant competitive ratio when the sample is uniform---if not, we show we can recover the same result by adding a provably minimal sample. Empirical results validate the theory and show that such algorithms can be used on temporal data to obtain strong results.

Standard approaches for global optimization of non-convex functions, such as branch-and-bound, maintain partition trees to systematically prune the domain. The tree size grows exponentially in the number of dimensions. We propose new sampling-based methods for non-convex optimization that adapts Monte Carlo Tree Search (MCTS) to improve efficiency. Instead of the standard use of visitation count in Upper Confidence Bounds, we utilize numerical overapproximations of the objective as an uncertainty metric, and also take into account of sampled estimates of first-order and second-order information. The Monte Carlo tree in our approach avoids the usual fixed combinatorial patterns in growing the tree, and aggressively zooms into the promising regions, while still balancing exploration and exploitation. We evaluate the proposed algorithms on high-dimensional non-convex optimization benchmarks against competitive baselines and analyze the effects of the hyper parameters.

Incomplete multi-view clustering has attracted much attention due to its ability to handle partial multi-view data. Recently, similarity-based methods have been developed to explore the complete relationship among incomplete multi-view data. Although widely applied to partial scenarios, most of the existing approaches are still faced with two limitations. Firstly, fusing similarities constructed individually on each view fails to yield a complete unified similarity. Moreover, incomplete similarity generation may lead to anomalous similarity values with column sum constraints, affecting the final clustering results. To solve the above challenging issues, we propose a Sample-level Cross-view Similarity Learning (SCSL) method for Incomplete Multi-view Clustering. Specifically, we project all samples to the same dimension and simultaneously construct a complete similarity matrix across views based on the inter-view sample relationship and the intra-view sample relationship. In addition, a simultaneously learning consensus representation ensures the validity of the projection, which further enhances the quality of the similarity matrix through the graph Laplacian regularization. Experimental results on six benchmark datasets demonstrate the ability of SCSL in processing incomplete multi-view clustering tasks. Our code is publicly available at https://github.com/Tracesource/SCSL.

We consider the problem of personalizing audio to maximize user experience. Briefly, we aim to find a filter h*, which applied to any music or speech, will maximize the user’s satisfaction. This is a black-box optimization problem since the user’s satisfaction function is unknown. Substantive work has been done on this topic where the key idea is to play audio samples to the user, each shaped by a different filter hi, and query the user for their satisfaction scores f(hi). A family of “surrogate” functions is then designed to fit these scores and the optimization method gradually refines these functions to arrive at the filter ˆh* that maximizes satisfaction. In certain applications, we observe that a second type of querying is possible where users can tell us the individual elements h*[j] of the optimal filter h*. Consider an analogy from cooking where the goal is to cook a recipe that maximizes user satisfaction. A user can be asked to score various cooked recipes (e.g., tofu fried rice) or to score individual ingredients (say, salt, sugar, rice, chicken, etc.). Given a budget of B queries, where a query can be of either type, our goal is to find the recipe that will maximize this user’s satisfaction. Our proposal builds on Sparse Gaussian Process Regression (GPR) and shows how a hybrid approach can outperform any one type of querying. Our results are validated through simulations and real world experiments, where volunteers gave feedback on music/speech audio and were able to achieve high satisfaction levels. We believe this idea of hybrid querying opens new problems in black-box optimization and solutions can benefit other applications beyond audio personalization.

The problem of sample complexity of online reinforcement learning is often studied in the literature without taking into account any partial knowledge about the system dynamics that could potentially accelerate the learning process. In this paper, we study the sample complexity of online Q-learning methods when some prior knowledge about the dynamics is available or can be learned efficiently. We focus on systems that evolve according to an additive disturbance model of the form S_{h+1} = ƒ(S_h, A_h) + W_h, where ƒ represents the underlying system dynamics, and W_h are unknown disturbances independent of states and actions. In the setting of finite episodic Markov decision processes with S states, A actions, and episode length H, we present an optimistic Q-learning algorithm that achieves Õ(Poly(H)√T) regret under perfect knowledge of ƒ, where T is the total number of interactions with the system. This is in contrast to the typical Õ(Poly(H)√SAT) regret for existing Q-learning methods. Further, if only a noisy estimate ƒ_hat of ƒ is available, our method can learn an approximately optimal policy in a number of samples that is independent of the cardinalities of state and action spaces. The sub-optimality gap depends on the approximation error ƒ_hat − ƒ, as well as the Lipschitz constant of the corresponding optimal value function. Our approach does not require modeling of the transition probabilities and enjoys the same memory complexity as model-free methods.

This paper addresses the problem of maintaining safety during training in Reinforcement Learning (RL), such that the safety constraint violations are bounded at any point during learning. As enforcing safety during training might severely limit the agent’s exploration, we propose here a new architecture that handles the trade-off between efficient progress and safety during exploration. As the exploration progresses, we update via Bayesian inference Dirichlet-Categorical models of the transition probabilities of the Markov decision process that describes the environment dynamics. We then propose a way to approximate moments of belief about the risk associated to the action selection policy. We demonstrate that this approach can be easily interleaved with RL and we present experimental results to showcase the performance of the overall architecture.

With the growing use of machine learning (ML) models in critical domains such as finance and healthcare, the need to offer recourse for those adversely affected by the decisions of ML models has become more important; individuals ought to be provided with recommendations on actions to take for improving their situation and thus receiving a favorable decision. Prior work on sequential algorithmic recourse---which recommends a series of changes---focuses on action feasibility and uses the proximity of feature changes to determine action costs. However, the uncertainties of feature changes and the risk of higher than average costs in recourse have not been considered. It is undesirable if a recourse could (with some probability) result in a worse situation from which recovery requires an extremely high cost. It is essential to incorporate risks when computing and evaluating recourse. We call the recourse computed with such risk considerations as Safe Algorithmic Recourse (SafeAR). The objective is to empower people to choose a recourse based on their risk tolerance. In this work, we discuss and show how existing recourse desiderata can fail to capture the risk of higher costs. We present a method to compute recourse policies that consider variability in cost and connect algorithmic recourse literature with risk-sensitive reinforcement learning. We also adopt measures "Value at Risk" and "Conditional Value at Risk" from the financial literature to summarize risk concisely. We apply our method to two real-world datasets and compare policies with different risk-aversion levels using risk measures and recourse desiderata (sparsity and proximity).

This paper studies safe Reinforcement Learning (safe RL) with linear function approximation and under hard instantaneous constraints where unsafe actions must be avoided at each step. Existing studies have considered safe RL with hard instantaneous constraints, but their approaches rely on several key assumptions: (i) the RL agent knows a safe action set for every state or knows a safe graph in which all the state-action-state triples are safe, and (ii) the constraint/cost functions are linear. In this paper, we consider safe RL with instantaneous hard constraints without assumption (i) and generalize (ii) to Reproducing Kernel Hilbert Space (RKHS). Our proposed algorithm, LSVI-AE, achieves O(√{d³H⁴K}) regret and O(H √{dK}) hard constraint violation when the cost function is linear and O(H?ₖ √{K}) hard constraint violation when the cost function belongs to RKHS. Here K is the learning horizon, H is the length of each episode, and ?ₖ is the information gain w.r.t the kernel used to approximate cost functions. Our results achieve the optimal dependency on the learning horizon K, matching the lower bound we provide in this paper and demonstrating the efficiency of LSVI-AE. Notably, the design of our approach encourages aggressive policy exploration, providing a unique perspective on safe RL with general cost functions and no prior knowledge of safe actions, which may be of independent interest.

Integrating complementary strengths of raw data and logical rules to improve the learning generalization has been recently shown promising and effective, e.g., abductive learning is one generic framework that can learn the perception model from data and reason between rules simultaneously. However, the performance would be seriously decreased when inaccurate logical rules appear, which may be even worse than baselines using only raw data. Efforts on this issue are highly desired while remain to be limited. This paper proposes a simple and effective safe abductive learning method to alleviate the harm caused by inaccurate rules. Unlike the existing methods which directly use all rules without correctness checks, it utilizes them selectively by constructing a graphical model with an adaptive reasoning process to prevent performance hazards. Theoretically, we show that induction and abduction are mutually beneficial, and can be rigorously justified from a classical maximum likelihood estimation perspective. Experiments on diverse tasks show that our method can tolerate at least twice as many inaccurate rules as accurate ones and achieve highly competitive performance while other methods can't. Moreover, the proposal can refine inaccurate rules and works well in extended weakly supervised scenarios.

Deep Multi-view Graph Clustering (DMGC) aims to partition instances into different groups using the graph information extracted from multi-view data. The mainstream framework of DMGC methods applies graph neural networks to embed structure information into the view-specific representations and fuse them for the consensus representation. However, on one hand, we find that the graph learned in advance is not ideal for clustering as it is constructed by original multi-view data and localized connecting. On the other hand, most existing methods learn the consensus representation in a late fusion manner, which fails to propagate the structure relations across multiple views. Inspired by the observations, we propose a Structure-adaptive Unified gRaph nEural network for multi-view clusteRing (SURER), which can jointly learn a heterogeneous multi-view unified graph and robust graph neural networks for multi-view clustering. Specifically, we first design a graph structure learning module to refine the original view-specific attribute graphs, which removes false edges and discovers the potential connection. According to the view-specific refined attribute graphs, we integrate them into a unified heterogeneous graph by linking the representations of the same sample from different views. Furthermore, we use the unified heterogeneous graph as the input of the graph neural network to learn the consensus representation for each instance, effectively integrating complementary information from various views. Extensive experiments on diverse datasets demonstrate the superior effectiveness of our method compared to other state-of-the-art approaches.

Offline-to-online reinforcement learning (RL) provides a promising solution to improving suboptimal offline pre-trained policies through online fine-tuning. However, one efficient method, unconstrained fine-tuning, often suffers from severe policy collapse due to excessive distribution shift. To ensure stability, existing methods retain offline constraints and employ additional techniques during fine-tuning, which hurts efficiency. In this work, we introduce a novel perspective: eliminating the policy collapse without imposing constraints. We observe that such policy collapse arises from the mismatch between unconstrained fine-tuning and the conventional RL training framework. To this end, we propose Stabilized Unconstrained Fine-tuning (SUF), a streamlined framework that benefits from the efficiency of unconstrained fine-tuning while ensuring stability by modifying the Update-To-Data ratio. With just a few lines of code adjustments, SUF demonstrates remarkable adaptability to diverse backbones and superior performance over state-of-the-art baselines.

Multi-camera depth estimation has recently garnered significant attention due to its substantial practical implications in the realm of autonomous driving. In this paper, we delve into the task of self-supervised multi-camera depth estimation and propose an innovative framework, STViT, featuring several noteworthy enhancements: 1) we propose a Spatial-Temporal Transformer to comprehensively exploit both local connectivity and the global context of image features, meanwhile learning enriched spatial-temporal cross-view correlations to recover 3D geometry. 2) to alleviate the severe effect of adverse conditions, e.g., rainy weather and nighttime driving, we introduce a GAN-based Adversarial Geometry Regularization Module (AGR) to further constrain the depth estimation with unpaired normal-condition depth maps and prevent the model from being incorrectly trained. Experiments on challenging autonomous driving datasets Nuscenes and DDAD show that our method achieves state-of-the-art performance.

Multi-task learning (MTL) has gained significant popularity in recommender systems as it enables simultaneous optimization of multiple objectives. A key challenge in MTL is negative transfer, but existing studies explored negative transfer on all samples, overlooking the inherent complexities within them. We split the samples according to the relative amount of positive feedback among tasks. Surprisingly, negative transfer still occurs in existing MTL methods on samples that receive comparable feedback across tasks. Existing work commonly employs a shared-embedding paradigm, limiting the ability of modeling diverse user preferences on different tasks. In this paper, we introduce a novel Shared and Task-specific EMbeddings (STEM) paradigm that aims to incorporate both shared and task-specific embeddings to effectively capture task-specific user preferences. Under this paradigm, we propose a simple model STEM-Net, which is equipped with an All Forward Task-specific Backward gating network to facilitate the learning of task-specific embeddings and direct knowledge transfer across tasks. Remarkably, STEM-Net demonstrates exceptional performance on comparable samples, achieving positive transfer. Comprehensive evaluation on three public MTL recommendation datasets demonstrates that STEM-Net outperforms state-of-the-art models by a substantial margin. Our code is released at https://github.com/LiangcaiSu/STEM.

Predicting future frames of a video is challenging because it is difficult to learn the uncertainty of the underlying factors influencing their contents. In this paper, we propose a novel video prediction model, which has infinite-dimensional latent variables over the spatio-temporal domain. Specifically, we first decompose the video motion and content information, then take a neural stochastic differential equation to predict the temporal motion information, and finally, an image diffusion model autoregressively generates the video frame by conditioning on the predicted motion feature and the previous frame. The better expressiveness and stronger stochasticity learning capability of our model lead to state-of-the-art video prediction performances. As well, our model is able to achieve temporal continuous prediction, i.e., predicting in an unsupervised way the future video frames with an arbitrarily high frame rate. Our code is available at https://github.com/XiYe20/STDiffProject.

Centralized Training with Decentralized Execution (CTDE) has been proven to be an effective paradigm in cooperative multi-agent reinforcement learning (MARL). One of the major challenges is credit assignment, which aims to credit agents by their contributions. They lack the functionality to model complicated relations of the delayed global reward in the temporal dimension and suffer from inefficiencies. To tackle this, we introduce Spatial-Temporal Attention with Shapley (STAS), a novel method that learns credit assignment in both temporal and spatial dimensions. It first decomposes the global return back to each time step, then utilizes the Shapley Value to redistribute the individual payoff from the decomposed global reward. To mitigate the computational complexity of the Shapley Value, we introduce an approximation of marginal contribution and utilize Monte Carlo sampling to estimate it. We evaluate our method on an Alice & Bob example and MPE environments across different scenarios. Our results demonstrate that our method effectively assigns spatial-temporal credit, outperforming all state-of-the-art baselines.

Information extraction tasks such as event extraction require an in-depth understanding of the output structure and sub-task dependencies. They heavily rely on task-specific training data in the form of (passage, target structure) pairs to obtain reasonable performance. However, obtaining such data through human annotation is costly, leading to a pressing need for low-resource information extraction approaches that require minimal human labeling for real-world applications. Fine-tuning supervised models with synthesized training data would be a generalizable method, but the existing data generation methods either still rely on large-scale ground-truth data or cannot be applied to complicated IE tasks due to their poor performance. To address these challenges, we propose STAR, a data generation method that leverages Large Language Models (LLMs) to synthesize data instances given limited seed demonstrations, thereby boosting low-resource information extraction performance. Our approach involves generating target structures (Y) followed by generating passages (X), all accomplished with the aid of LLMs. We design fine-grained step-by-step instructions to obtain the initial data instances. We further reduce errors and improve data quality through self-reflection error identification and self-refinement with iterative revision. Our experiments show that the data generated by STAR significantly improve the performance of low-resource event extraction and relation extraction tasks, even surpassing the effectiveness of human-curated data. Human assessment of the data quality shows STAR-generated data exhibit higher passage quality and better align with the task definitions compared with the human-curated data.