NeurIPS 2024 Papers

We propose an unsupervised adaptation framework, Self-TAught Recognizer (STAR), which leverages unlabeled data to enhance the robustness of automatic speech recognition (ASR) systems in diverse target domains, such as noise and accents. STAR is developed for prevalent speech foundation models based on Transformer-related architecture with auto-regressive decoding (e.g., Whisper, Canary). Specifically, we propose a novel indicator that empirically integrates step-wise information during decoding to assess the token-level quality of pseudo labels without ground truth, thereby guiding model updates for effective unsupervised adaptation. Experimental results show that STAR achieves an average of 13.5% relative reduction in word error rate across 14 target domains, and it sometimes even approaches the upper-bound performance of supervised adaptation. Surprisingly, we also observe that STAR prevents the adapted model from the common catastrophic forgetting problem without recalling source-domain data. Furthermore, STAR exhibits high data efficiency that only requires less than one-hour unlabeled data, and seamless generality to alternative large speech models and speech translation tasks. Our code aims to open source to the research communities.

Deep learning-based radar detection technology is receiving increasing attention in areas such as autonomous driving, UAV surveillance, and marine monitoring. Among recent efforts, PeakConv (PKC) provides a solution that can retain the peak response characteristics of radar signals and play the characteristics of deep convolution, thereby improving the effect of radar semantic segmentation (RSS). However, due to the use of a pre-set fixed peak receptive field sampling rule, PKC still has limitations in dealing with problems such as inconsistency of target frequency domain response broadening, non-homogeneous and time-varying characteristic of noise/clutter distribution. Therefore, this paper proposes an idea of adaptive peak receptive field, and upgrades PKC to AdaPKC based on this idea. Beyond that, a novel fine-tuning technology to further boost the performance of AdaPKC-based RSS networks is presented. Through experimental verification using various real-measured radar data (including publicly available low-cost millimeter-wave radar dataset for autonomous driving and self-collected Ku-band surveillance radar dataset), we found that the performance of AdaPKC-based models surpasses other SoTA methods in RSS tasks. The code is available at https://github.com/lihua199710/AdaPKC.

The comic domain is rapidly advancing with the development of single-page analysis and synthesis models. However, evaluation metrics and datasets lag behind, often limited to small-scale or single-style test sets. We introduce a novel benchmark, CoMix, designed to evaluate the multi-task capabilities of models in comic analysis. Unlike existing benchmarks that focus on isolated tasks such as object detection or text recognition, CoMix addresses a broader range of tasks including object detection, speaker identification, character re-identification, reading order, and multi-modal reasoning tasks like character naming and dialogue generation. Our benchmark comprises three existing datasets with expanded annotations to support multi-task evaluation. To mitigate the over-representation of manga-style data, we have incorporated a new dataset of carefully selected American comic-style books, thereby enriching the diversity of comic styles. CoMix is designed to assess pre-trained models in zero-shot and limited fine-tuning settings, probing their transfer capabilities across different comic styles and tasks. The validation split of the benchmark is publicly available for research purposes, and an evaluation server for the held-out test split is also provided. Comparative results between human performance and state-of-the-art models reveal a significant performance gap, highlighting substantial opportunities for advancements in comic understanding. The dataset, baseline models, and code are accessible at https://github.com/emanuelevivoli/CoMix-dataset. This initiative sets a new standard for comprehensive comic analysis, providing the community with a common benchmark for evaluation on a large and varied set.

Most existing out-of-distribution (OOD) detection benchmarks classify samples with novel labels as the OOD data. However, some marginal OOD samples actually have close semantic contents to the in-distribution (ID) sample, which makes determining the OOD sample a Sorites Paradox. In this paper, we construct a benchmark named Incremental Shift OOD (IS-OOD) to address the issue, in which we divide the test samples into subsets with different semantic and covariate shift degrees relative to the ID dataset. The data division is achieved through a shift measuring method based on our proposed Language Aligned Image feature Decomposition (LAID). Moreover, we construct a Synthetic Incremental Shift (Syn-IS) dataset that contains high-quality generated images with more diverse covariate contents to complement the IS-OOD benchmark. We evaluate current OOD detection methods on our benchmark and find several important insights: (1) The performance of most OOD detection methods significantly improves as the semantic shift increases; (2) Some methods like GradNorm may have different OOD detection mechanisms as they rely less on semantic shifts to make decisions; (3) Excessive covariate shifts in the image are also likely to be considered as OOD for some methods. Our code and data are released in https://github.com/qqwsad5/IS-OOD.

In this work, we observe that model trained on vast general images via masking strategy, has been naturally embedded with their distribution knowledge, thus spontaneously attains the underlying potential for strong image denoising. Based on this observation, we propose a novel zero-shot denoising paradigm, i.e., $\textbf{M}$asked $\textbf{P}$re-train then $\textbf{I}$terative fill ($\textbf{MPI}$). MPI first trains model via masking and then employs pre-trained weight for high-quality zero-shot image denoising on a single noisy image. Concretely, MPI comprises two key procedures: $\textbf{1) Masked Pre-training}$ involves training model to reconstruct massive natural images with random masking for generalizable representations, gathering the potential for valid zero-shot denoising on images with varying noise degradation and even in distinct image types. $\textbf{2) Iterative filling}$ exploits pre-trained knowledge for effective zero-shot denoising. It iteratively optimizes the image by leveraging pre-trained weights, focusing on alternate reconstruction of different image parts, and gradually assembles fully denoised image within limited number of iterations. Comprehensive experiments across various noisy scenarios underscore the notable advances of MPI over previous approaches with a marked reduction in inference time.

Since the inception of the classicalist vs. connectionist debate, it has been argued that the ability to systematically combine symbol-like entities into compositional representations is crucial for human intelligence. In connectionist systems, the field of disentanglement has gained prominence for its ability to produce explicitly compositional representations; however, it relies on a fundamentally *symbolic, concatenative* representation of compositional structure that clashes with the *continuous, distributed* foundations of deep learning. To resolve this tension, we extend Smolensky's Tensor Product Representation (TPR) and introduce *Soft TPR*, a representational form that encodes compositional structure in an inherently *distributed, flexible* manner, along with *Soft TPR Autoencoder*, a theoretically-principled architecture designed specifically to learn Soft TPRs. Comprehensive evaluations in the visual representation learning domain demonstrate that the Soft TPR framework consistently outperforms conventional disentanglement alternatives -- achieving state-of-the-art disentanglement, boosting representation learner convergence, and delivering superior sample efficiency and low-sample regime performance in downstream tasks. These findings highlight the promise of a *distributed* and *flexible* approach to representing compositional structure by potentially enhancing alignment with the core principles of deep learning over the conventional symbolic approach.

Global routing plays a critical role in modern chip design. The routing paths generated by global routers often form a rectilinear Steiner tree (RST). Recent advances from the machine learning community have shown the power of learning-based route generation; however, the yielded routing paths by the existing approaches often suffer from considerable overflow, thus greatly hindering their application in practice. We propose NeuralSteiner, an accurate approach to overflow-avoiding global routing in chip design. The key idea of NeuralSteiner approach is to learn Steiner trees: we first predict the locations of highly likely Steiner points by adopting a neural network considering full-net spatial and overflow information, then select appropriate points by running a graph-based post-processing algorithm, and finally connect these points with the input pins to yield overflow-avoiding RSTs. NeuralSteiner offers two advantages over previous learning-based models. First, by using the learning scheme, NeuralSteiner ensures the connectivity of generated routes while significantly reducing congestion. Second, NeuralSteiner can effectively scale to large nets and transfer to unseen chip designs without any modifications or fine-tuning. Extensive experiments over public large-scale benchmarks reveal that, compared with the state-of-the-art deep generative methods, NeuralSteiner achieves up to a 99.8\% reduction in overflow while speeding up the generation and maintaining a slight wirelength loss within only 1.8\%.

The effect of regularizers such as weight decay when training deep neural networks is not well understood. We study the influence of weight decay as well as $L2$-regularization when training neural network models in which parameter matrices interact multiplicatively. This combination is of particular interest as this parametrization is common in attention layers, the workhorse of transformers. Here, key-query, as well as value-projection parameter matrices, are multiplied directly with each other: $W_K^TW_Q$ and $PW_V$. We extend previous results and show on one hand that any local minimum of a $L2$-regularized loss of the form $L(AB^\top) + \lambda (\|A\|^2 + \|B\|^2)$ coincides with a minimum of the nuclear norm-regularized loss $L(AB^\top) + \lambda\|AB^\top\|_*$, and on the other hand that the 2 losses become identical exponentially quickly during training. We thus complement existing works linking $L2$-regularization with low-rank regularization, and in particular, explain why such regularization on the matrix product affects early stages of training. Based on these theoretical insights, we verify empirically that the key-query and value-projection matrix products $W_K^TW_Q, PW_V$ within attention layers, when optimized with weight decay, as usually done in vision tasks and language modelling, indeed induce a significant reduction in the rank of $W_K^TW_Q$ and $PW_V$, even in fully online training. We find that, in accordance with existing work, inducing low rank in attention matrix products can damage language model performance, and observe advantages when decoupling weight decay in attention layers from the rest of the parameters.

We explore the potential of self-play training for large language models (LLMs) in a two-player adversarial language game called Adversarial Taboo. In this game, an attacker and a defender communicate around a target word only visible to the attacker. The attacker aims to induce the defender to speak the target word unconsciously, while the defender tries to infer the target word from the attacker's utterances. To win the game, both players must have sufficient knowledge about the target word and high-level reasoning ability to infer and express in this information-reserved conversation. Hence, we are curious about whether LLMs' reasoning ability can be further enhanced by Self-Playing this Adversarial language Game (SPAG). With this goal, we select several open-source LLMs and let each act as the attacker and play with a copy of itself as the defender on an extensive range of target words. Through reinforcement learning on the game outcomes, we observe that the LLMs' performances uniformly improve on a broad range of reasoning benchmarks. Furthermore, iteratively adopting this self-play process can continuously promote LLMs' reasoning abilities. The code is available at https://github.com/Linear95/SPAG.

Automatically generating novel and interesting games is a complex task. Challenges include representing game rules in a computationally workable form, searching through the large space of potential games under most such representations, and accurately evaluating the originality and quality of previously unseen games. Prior work in automated game generation has largely focused on relatively restricted rule representations and relied on domain-specific heuristics. In this work, we explore the generation of novel games in the comparatively expansive Ludii game description language, which encodes the rules of over 1000 board games in a variety of styles and modes of play. We draw inspiration from recent advances in large language models and evolutionary computation in order to train a model that intelligently mutates and recombines games and mechanics expressed as code. We demonstrate both quantitatively and qualitatively that our approach is capable of generating new and interesting games, including in regions of the potential rules space not covered by existing games in the Ludii dataset.

Diffusion models have obtained substantial progress in image-to-video generation. However, in this paper, we find that these models tend to generate videos with less motion than expected. We attribute this to the issue called conditional image leakage, where the image-to-video diffusion models (I2V-DMs) tend to over-rely on the conditional image at large time steps. We further address this challenge from both inference and training aspects. First, we propose to start the generation process from an earlier time step to avoid the unreliable large-time steps of I2V-DMs, as well as an initial noise distribution with optimal analytic expressions (Analytic-Init) by minimizing the KL divergence between it and the actual marginal distribution to bridge the training-inference gap. Second, we design a time-dependent noise distribution (TimeNoise) for the conditional image during training, applying higher noise levels at larger time steps to disrupt it and reduce the model's dependency on it. We validate these general strategies on various I2V-DMs on our collected open-domain image benchmark and the UCF101 dataset. Extensive results show that our methods outperform baselines by producing higher motion scores with lower errors while maintaining image alignment and temporal consistency, thereby yielding superior overall performance and enabling more accurate motion control. The project page: \url{https://cond-image-leak.github.io/}.

Cross-domain few-shot learning (CDFSL) is proposed to transfer knowledge from large-scale source-domain datasets to downstream target-domain datasets with only a few training samples. However, Vision Transformer (ViT), as a strong backbone network to achieve many top performances, is still under-explored in the CDFSL task in its transferability against large domain gaps. In this paper, we find an interesting phenomenon of ViT in the CDFSL task: by simply multiplying a temperature (even as small as 0) to the attention in ViT blocks, the target-domain performance consistently increases, even though the attention map is downgraded to a uniform map. In this paper, we delve into this phenomenon for an interpretation. Through experiments, we interpret this phenomenon as a remedy for the ineffective target-domain attention caused by the query-key attention mechanism under large domain gaps. Based on it, we further propose a simple but effective method for the CDFSL task to boost ViT's transferability by resisting the learning of query-key parameters and encouraging that of non-query-key ones. Experiments on four CDFSL datasets validate the rationale of our interpretation and method, showing we can consistently outperform state-of-the-art methods. Our codes are available at https://github.com/Zoilsen/Attn_Temp_CDFSL.

Iteratively improving and repairing source code with large language models (LLMs), known as refinement, has emerged as a popular way of generating programs that would be too complex to construct in one shot. Given a bank of test cases, together with a candidate program, an LLM can improve that program by being prompted with failed test cases. But it remains an open question how to best iteratively refine code, with prior work employing simple greedy or breadth-first strategies. We show here that refinement exposes an explore-exploit tradeoff: exploit by refining the program that passes the most test cases, or explore by refining a lesser considered program. We frame this as an arm-acquiring bandit problem, which we solve with Thompson Sampling. The resulting LLM-based program synthesis algorithm is broadly applicable: Across loop invariant synthesis, visual reasoning puzzles, and competition programming problems, we find that our new method can solve more problems using fewer language model calls.

High-resolution images offer more information about scenes that can improve model accuracy. However, the dominant model architecture in computer vision, the vision transformer (ViT), cannot effectively leverage larger images without finetuning — ViTs poorly extrapolate to more patches at test time, although transformers offer sequence length flexibility. We attribute this shortcoming to the current patch position encoding methods, which create a distribution shift when extrapolating. We propose a drop-in replacement for the position encoding of plain ViTs that restricts attention heads to fixed fields of view, pointed in different directions, using 2D attention masks. Our novel method, called LookHere, provides translation-equivariance, ensures attention head diversity, and limits the distribution shift that attention heads face when extrapolating. We demonstrate that LookHere improves performance on classification (avg. 1.6%), against adversarial attack (avg. 5.4%), and decreases calibration error (avg. 1.5%) — on ImageNet without extrapolation. With extrapolation, LookHere outperforms the current SoTA position encoding method, 2D-RoPE, by 21.7% on ImageNet when trained at $224^2$ px and tested at $1024^2$ px. Additionally, we release a high-resolution test set to improve the evaluation of high-resolution image classifiers, called ImageNet-HR.

This work considers the problem of sampling from a probability distribution known up to a normalization constant while satisfying a set of statistical constraints specified by the expected values of general nonlinear functions. This problem finds applications in, e.g., Bayesian inference, where it can constrain moments to evaluate counterfactual scenarios or enforce desiderata such as prediction fairness. Methods developed to handle support constraints, such as those based on mirror maps, barriers, and penalties, are not suited for this task. This work therefore relies on gradient descent-ascent dynamics in Wasserstein space to put forward a discrete-time primal-dual Langevin Monte Carlo algorithm (PD-LMC) that simultaneously constrains the target distribution and samples from it. We analyze the convergence of PD-LMC under standard assumptions on the target distribution and constraints, namely (strong) convexity and log-Sobolev inequalities. To do so, we bring classical optimization arguments for saddle-point algorithms to the geometry of Wasserstein space. We illustrate the relevance and effectiveness of PD-LMC in several applications.

Partial observability of the underlying states generally presents significant challenges for reinforcement learning (RL). In practice, certain *privileged information* , e.g., the access to states from simulators, has been exploited in training and achieved prominent empirical successes. To better understand the benefits of privileged information, we revisit and examine several simple and practically used paradigms in this setting, with both computation and sample efficiency analyses. Specifically, we first formalize the empirical paradigm of *expert distillation* (also known as *teacher-student* learning), demonstrating its pitfall in finding near-optimal policies. We then identify a condition of the partially observable environment, the deterministic filter condition, under which expert distillation achieves sample and computational complexities that are *both* polynomial. Furthermore, we investigate another successful empirical paradigm of *asymmetric actor-critic*, and focus on the more challenging setting of observable partially observable Markov decision processes. We develop a belief-weighted optimistic asymmetric actor-critic algorithm with polynomial sample and quasi-polynomial computational complexities, where one key component is a new provable oracle for learning belief states that preserve *filter stability* under a misspecified model, which may be of independent interest. Finally, we also investigate the provable efficiency of partially observable multi-agent RL (MARL) with privileged information. We develop algorithms with the feature of centralized-training-with-decentralized-execution, a popular framework in empirical MARL, with polynomial sample and (quasi-)polynomial computational complexity in both paradigms above. Compared with a few recent related theoretical studies, our focus is on understanding practically inspired algorithmic paradigms, without computationally intractable oracles.

While imitation learning requires access to high-quality data, offline reinforcement learning (RL) should, in principle, perform similarly or better with substantially lower data quality by using a value function. However, current results indicate that offline RL often performs worse than imitation learning, and it is often unclear what holds back the performance of offline RL. Motivated by this observation, we aim to understand the bottlenecks in current offline RL algorithms. While poor performance of offline RL is typically attributed to an imperfect value function, we ask: *is the main bottleneck of offline RL indeed in learning the value function, or something else?* To answer this question, we perform a systematic empirical study of (1) value learning, (2) policy extraction, and (3) policy generalization in offline RL problems, analyzing how these components affect performance. We make two surprising observations. First, we find that the choice of a policy extraction algorithm significantly affects the performance and scalability of offline RL, often more so than the value learning objective. For instance, we show that common value-weighted behavioral cloning objectives (e.g., AWR) do not fully leverage the learned value function, and switching to behavior-constrained policy gradient objectives (e.g., DDPG+BC) often leads to substantial improvements in performance and scalability. Second, we find that a big barrier to improving offline RL performance is often imperfect policy generalization on test-time states out of the support of the training data, rather than policy learning on in-distribution states. We then show that the use of suboptimal but high-coverage data or test-time policy training techniques can address this generalization issue in practice. Specifically, we propose two simple test-time policy improvement methods and show that these methods lead to better performance.

Graph Neural Networks (GNNs) have demonstrated remarkable performance across a spectrum of graph-related tasks, however concerns persist regarding their vulnerability to adversarial perturbations. While prevailing defense strategies focus primarily on pre-processing techniques and adaptive message-passing schemes, this study delves into an under-explored dimension: the impact of weight initialization and associated hyper-parameters, such as training epochs, on a model’s robustness. We introduce a theoretical framework bridging the connection between initialization strategies and a network's resilience to adversarial perturbations. Our analysis reveals a direct relationship between initial weights, number of training epochs and the model’s vulnerability, offering new insights into adversarial robustness beyond conventional defense mechanisms. While our primary focus is on GNNs, we extend our theoretical framework, providing a general upper-bound applicable to Deep Neural Networks. Extensive experiments, spanning diverse models and real-world datasets subjected to various adversarial attacks, validate our findings. We illustrate that selecting appropriate initialization not only ensures performance on clean datasets but also enhances model robustness against adversarial perturbations, with observed gaps of up to 50\% compared to alternative initialization approaches.

With the rapid development of deep learning, the increasing complexity and scale of parameters make training a new model increasingly resource-intensive. In this paper, we start from the classic convolutional neural network (CNN) and explore a paradigm that does not require training to obtain new models. Similar to the birth of CNN inspired by receptive fields in the biological visual system, we draw inspiration from the information subsystem pathways in the biological visual system and propose Model Disassembling and Assembling (MDA). During model disassembling, we introduce the concept of relative contribution and propose a component locating technique to extract task-aware components from trained CNN classifiers. For model assembling, we present the alignment padding strategy and parameter scaling strategy to construct a new model tailored for a specific task, utilizing the disassembled task-aware components. The entire process is akin to playing with LEGO bricks, enabling arbitrary assembly of new models, and providing a novel perspective for model creation and reuse. Extensive experiments showcase that task-aware components disassembled from CNN classifiers or new models assembled using these components closely match or even surpass the performance of the baseline, demonstrating its promising results for model reuse. Furthermore, MDA exhibits diverse potential applications, with comprehensive experiments exploring model decision route analysis, model compression, knowledge distillation, and more.