NeurIPS 2023 Papers

Acoustic matching aims to re-synthesize an audio clip to sound as if it were recorded in a target acoustic environment. Existing methods assume access to paired training data, where the audio is observed in both source and target environments, but this limits the diversity of training data or requires the use of simulated data or heuristics to create paired samples. We propose a self-supervised approach to visual acoustic matching where training samples include only the target scene image and audio---without acoustically mismatched source audio for reference. Our approach jointly learns to disentangle room acoustics and re-synthesize audio into the target environment, via a conditional GAN framework and a novel metric that quantifies the level of residual acoustic information in the de-biased audio. Training with either in-the-wild web data or simulated data, we demonstrate it outperforms the state-of-the-art on multiple challenging datasets and a wide variety of real-world audio and environments.

Self-supervised training methods for transformers have demonstrated remarkable performance across various domains. Previous transformer-based models, such as masked autoencoders (MAE), typically utilize a single normalization layer for both the [CLS] symbol and the tokens. We propose in this paper a simple modification that employs separate normalization layers for the tokens and the [CLS] symbol to better capture their distinct characteristics and enhance downstream task performance. Our method aims to alleviate the potential negative effects of using the same normalization statistics for both token types, which may not be optimally aligned with their individual roles. We empirically show that by utilizing a separate normalization layer, the [CLS] embeddings can better encode the global contextual information and are distributed more uniformly in its anisotropic space. When replacing the conventional normalization layer with the two separate layers, we observe an average 2.7% performance improvement over the image, natural language, and graph domains.

Semi-Supervised Domain Generalization (SSDG) aims to learn a model that is generalizable to an unseen target domain with only a few labels, and most existing SSDG methods assume that unlabeled training and testing samples are all known classes. However, a more realistic scenario is that known classes may be mixed with some unknown classes in unlabeled training and testing data. To deal with such a scenario, we propose the Class-Wise Adaptive Exploration and Exploitation (CWAEE) method. In particular, we explore unlabeled training data by using one-vs-rest classifiers and class-wise adaptive thresholds to detect known and unknown classes, and exploit them by adopting consistency regularization on augmented samples based on Fourier Transformation to improve the unseen domain generalization. The experiments conducted on real-world datasets verify the effectiveness and superiority of our method.

Data imbalance and open-ended distribution are two intrinsic characteristics of the real visual world. Though encouraging progress has been made in tackling each challenge separately, few works dedicated to combining them towards real-world scenarios. While several previous works have focused on classifying close-set samples and detecting open-set samples during testing, it's still essential to be able to classify unknown subjects as human beings. In this paper, we formally define a more realistic task as distribution-agnostic generalized category discovery (DA-GCD): generating fine-grained predictions for both close- and open-set classes in a long-tailed open-world setting. To tackle the challenging problem, we propose a Self-**Ba**lanced **Co**-Advice co**n**trastive framework (BaCon), which consists of a contrastive-learning branch and a pseudo-labeling branch, working collaboratively to provide interactive supervision to resolve the DA-GCD task. In particular, the contrastive-learning branch provides reliable distribution estimation to regularize the predictions of the pseudo-labeling branch, which in turn guides contrastive learning through self-balanced knowledge transfer and a proposed novel contrastive loss. We compare BaCon with state-of-the-art methods from two closely related fields: imbalanced semi-supervised learning and generalized category discovery. The effectiveness of BaCon is demonstrated with superior performance over all baselines and comprehensive analysis across various datasets. Our code is publicly available.

Recently, a new line of works has emerged to understand and improve self-attention in Transformers by treating it as a kernel machine. However, existing works apply the methods for symmetric kernels to the asymmetric self-attention, resulting in a nontrivial gap between the analytical understanding and numerical implementation. In this paper, we provide a new perspective to represent and optimize self-attention through asymmetric Kernel Singular Value Decomposition (KSVD), which is also motivated by the low-rank property of self-attention normally observed in deep layers. Through asymmetric KSVD, i) a primal-dual representation of self-attention is formulated, where the optimization objective is cast to maximize the projection variances in the attention outputs; ii) a novel attention mechanism, i.e., Primal-Attention, is proposed via the primal representation of KSVD, avoiding explicit computation of the kernel matrix in the dual; iii) with KKT conditions, we prove that the stationary solution to the KSVD optimization in Primal-Attention yields a zero-value objective. In this manner, KSVD optimization can be implemented by simply minimizing a regularization loss, so that low-rank property is promoted without extra decomposition. Numerical experiments show state-of-the-art performance of our Primal-Attention with improved efficiency. Moreover, we demonstrate that the deployed KSVD optimization regularizes Primal-Attention with a sharper singular value decay than that of the canonical self-attention, further verifying the great potential of our method. To the best of our knowledge, this is the first work that provides a primal-dual representation for the asymmetric kernel in self-attention and successfully applies it to modelling and optimization.

Image-based virtual try-on tasks remain challenging, primarily due to inherent complexities associated with non-rigid garment deformation modeling and strong feature entanglement of clothing within human body. Recent groundbreaking formulations, such as in-painting, cycle consistency, and knowledge distillation, have facilitated self-supervised generation of try-on images. However, these paradigms necessitate the disentanglement of garment features within human body features through auxiliary tasks, such as leveraging 'teacher knowledge' and dual generators. The potential presence of irresponsible prior knowledge in the auxiliary task can serve as a significant bottleneck for the main generator (e.g., 'student model') in the downstream task. Moreover, existing garment deformation methods lack the ability to perceive the correlation between the garment and the human body in the real world, leading to unrealistic alignment effects. To tackle these limitations, we present a new parser-free virtual try-on network based on unified self-cycle consistency (USC-PFN), which enables robust translation between different garments using just a single generator, faithfully replicating non-rigid geometric deformation of garments in real-life scenarios. Specifically, we first propose a self-cycle consistency architecture with a circular mode. It utilizes real unpaired garment-person images exclusively as input for training, effectively eliminating the impact of irresponsible prior knowledge at the model input end. Additionally, we formulate a Markov Random Field to simulate a more natural and realistic garment deformation. Furthermore, USC-PFN can leverage a general generator for self-supervised cycle training. Experiments demonstrate that our method achieves state-of-the-art performance on a popular virtual try-on benchmark.

Video self-supervised learning (VSSL) has made significant progress in recent years. However, the exact behavior and dynamics of these models under different forms of distribution shift are not yet known. In this paper, we comprehensively study the behavior of six popular self-supervised methods (v-SimCLR, v-MoCo, v-BYOL, v-SimSiam, v-DINO, v-MAE) in response to various forms of natural distribution shift, i.e., (i) context shift, (ii) viewpoint shift, (iii) actor shift, (iv) source shift, (v) generalizability to unknown classes (zero-shot), and (vi) open-set recognition. To perform this extensive study, we carefully craft a test bed consisting of 17 in-distribution and out-of-distribution benchmark pairs using available public datasets and a series of evaluation protocols to stress-test the different methods under the intended shifts. Our study uncovers a series of intriguing findings and interesting behaviors of VSSL methods. For instance, we observe that while video models generally struggle with context shifts, v-MAE and supervised learning exhibit more robustness. Moreover, our study shows that v-MAE is a strong temporal learner, whereas contrastive methods, v-SimCLR and v-MoCo, exhibit strong performances against viewpoint shifts. When studying the notion of open-set recognition, we notice a trade-off between closed-set and open-set recognition performance if the pretrained VSSL encoders are used without finetuning. We hope that our work will contribute to the development of robust video representation learning frameworks for various real-world scenarios. The project page and code are available at: https://pritamqu.github.io/OOD-VSSL.

Reconstruction of 3D neural fields from posed images has emerged as a promising method for self-supervised representation learning. The key challenge preventing the deployment of these 3D scene learners on large-scale video data is their dependence on precise camera poses from structure-from-motion, which is prohibitively expensive to run at scale. We propose a method that jointly reconstructs camera poses and 3D neural scene representations online and in a single forward pass. We estimate poses by first lifting frame-to-frame optical flow to 3D scene flow via differentiable rendering, preserving locality and shift-equivariance of the image processing backbone. SE(3) camera pose estimation is then performed via a weighted least-squares fit to the scene flow field. This formulation enables us to jointly supervise pose estimation and a generalizable neural scene representation via re-rendering the input video, and thus, train end-to-end and fully self-supervised on real-world video datasets. We demonstrate that our method performs robustly on diverse, real-world video, notably on sequences traditionally challenging to optimization-based pose estimation techniques.

In a wide range of multimodal tasks, contrastive learning has become a particularly appealing approach since it can successfully learn representations from abundant unlabeled data with only pairing information (e.g., image-caption or video-audio pairs). Underpinning these approaches is the assumption of multi-view redundancy - that shared information between modalities is necessary and sufficient for downstream tasks. However, in many real-world settings, task-relevant information is also contained in modality-unique regions: information that is only present in one modality but still relevant to the task. How can we learn self-supervised multimodal representations to capture both shared and unique information relevant to downstream tasks? This paper proposes FactorCL, a new multimodal representation learning method to go beyond multi-view redundancy. FactorCL is built from three new contributions: (1) factorizing task-relevant information into shared and unique representations, (2) capturing task-relevant information via maximizing MI lower bounds and removing task-irrelevant information via minimizing MI upper bounds, and (3) multimodal data augmentations to approximate task relevance without labels. On large-scale real-world datasets, FactorCL captures both shared and unique information and achieves state-of-the-art results on six benchmarks.

Viewing Transformers as interacting particle systems, we describe the geometry of learned representations when the weights are not time-dependent. We show that particles, representing tokens, tend to cluster toward particular limiting objects as time tends to infinity. Using techniques from dynamical systems and partial differential equations, we show that type of limiting object that emerges depends on the spectrum of the value matrix. Additionally, in the one-dimensional case we prove that the self-attention matrix converges to a low-rank Boolean matrix. The combination of these results mathematically confirms the empirical observation made by Vaswani et al. [ VSP`17 ] that leaders appear in a sequence of tokens when processed by Transformers.

Machine learning for differential equations paves the way for computationally efficient alternatives to numerical solvers, with potentially broad impacts in science and engineering. Though current algorithms typically require simulated training data tailored to a given setting, one may instead wish to learn useful information from heterogeneous sources, or from real dynamical systems observations that are messy or incomplete. In this work, we learn general-purpose representations of PDEs from heterogeneous data by implementing joint embedding methods for self-supervised learning (SSL), a framework for unsupervised representation learning that has had notable success in computer vision. Our representation outperforms baseline approaches to invariant tasks, such as regressing the coefficients of a PDE, while also improving the time-stepping performance of neural solvers. We hope that our proposed methodology will prove useful in the eventual development of general-purpose foundation models for PDEs.

Open-world semi-supervised learning aims at inferring both known and novel classes in unlabeled data, by harnessing prior knowledge from a labeled set with known classes. Despite its importance, there is a lack of theoretical foundations for this problem. This paper bridges the gap by formalizing a graph-theoretic framework tailored for the open-world setting, where the clustering can be theoretically characterized by graph factorization. Our graph-theoretic framework illuminates practical algorithms and provides guarantees. In particular, based on our graph formulation, we apply the algorithm called Spectral Open-world Representation Learning (SORL), and show that minimizing our loss is equivalent to performing spectral decomposition on the graph. Such equivalence allows us to derive a provable error bound on the clustering performance for both known and novel classes, and analyze rigorously when labeled data helps. Empirically, SORL can match or outperform several strong baselines on common benchmark datasets, which is appealing for practical usage while enjoying theoretical guarantees.

Labeling training data is a critical and expensive step in producing high accuracy ML models, whether training from scratch or fine-tuning. To make labeling more efficient, two major approaches are programmatic weak supervision (WS) and semi-supervised learning (SSL). More recent works have either explicitly or implicitly used techniques at their intersection, but in various complex and ad hoc ways. In this work, we define a simple, modular design space to study the use of SSL techniques for WS more systematically. Surprisingly, we find that fairly simple methods from our design space match the performance of more complex state-of-the-art methods, averaging a 3 p.p. increase in accuracy/F1-score across 8 standard WS benchmarks. Further, we provide practical guidance on when different components are worth their added complexity and training costs. Contrary to current understanding, we find using SSL is not necessary to obtain the best performance on most WS benchmarks but is more effective when: (1) end models are smaller, and (2) WS provides labels for only a small portion of training examples.

Semi-supervised object detection is crucial for 3D scene understanding, efficiently addressing the limitation of acquiring large-scale 3D bounding box annotations. Existing methods typically employ a teacher-student framework with pseudo-labeling to leverage unlabeled point clouds. However, producing reliable pseudo-labels in a diverse 3D space still remains challenging. In this work, we propose Diffusion-SS3D, a new perspective of enhancing the quality of pseudo-labels via the diffusion model for semi-supervised 3D object detection. Specifically, we include noises to produce corrupted 3D object size and class label distributions, and then utilize the diffusion model as a denoising process to obtain bounding box outputs. Moreover, we integrate the diffusion model into the teacher-student framework, so that the denoised bounding boxes can be used to improve pseudo-label generation, as well as the entire semi-supervised learning process. We conduct experiments on the ScanNet and SUN RGB-D benchmark datasets to demonstrate that our approach achieves state-of-the-art performance against existing methods. We also present extensive analysis to understand how our diffusion model design affects performance in semi-supervised learning. The source code will be available at https://github.com/luluho1208/Diffusion-SS3D.

Self-supervised learning is introduced to train graph neural networks (GNNs) by employing propagation-based GNNs designed for semi-supervised learning tasks. Unfortunately, this common choice tends to cause two serious issues. Firstly, global parameters cause the model lack the ability to capture the local property. Secondly, it is difficult to handle networks beyond homophily without label information. This paper tends to break through the common choice of employing propagation-based GNNs, which aggregate representations of nodes belonging to different classes and tend to lose discriminative information. If the propagation in each ego-network is just between the nodes from the same class, the obtained representation matrix should follow the low-rank characteristic. To meet this requirement, this paper proposes the Low-Rank Decomposition-based GNNs (LRD-GNN-Matrix) by employing Low-Rank Decomposition to the attribute matrix. Furthermore, to incorporate long-distance information, Low-Rank Tensor Decomposition-based GNN (LRD-GNN-Tensor) is proposed by constructing the node attribute tensor from selected similar ego-networks and performing Low-Rank Tensor Decomposition. The employed tensor nuclear norm facilitates the capture of the long-distance relationship between original and selected similar ego-networks. Extensive experiments demonstrate the superior performance and the robustness of LRD-GNNs.

In self-supervised representation learning, Siamese networks are a natural architecture for learning transformation-invariance by bringing representations of positive pairs closer together. But it is prone to collapse into a degenerate solution. To address the issue, in contrastive learning, a contrastive loss is used to prevent collapse by moving representations of negative pairs away from each other. But it is known that algorithms with negative sampling are not robust to a reduction in the number of negative samples. So, on the other hand, there are algorithms that do not use negative pairs. Many positive-only algorithms adopt asymmetric network architecture consisting of source and target encoders as a key factor in coping with collapse. By exploiting the asymmetric architecture, we introduce a methodology to implicitly incorporate the idea of contrastive learning. As its implementation, we present a novel method guided stop-gradient. We apply our method to benchmark algorithms SimSiam and BYOL and show that our method stabilizes training and boosts performance. We also show that the algorithms with our method work well with small batch sizes and do not collapse even when there is no predictor. The code is available in the supplementary material.

Diffusion model based language-guided image editing has achieved great success recently. However, existing state-of-the-art diffusion models struggle with rendering correct text and text style during generation. To tackle this problem, we propose a universal self-supervised text editing diffusion model (DiffUTE), which aims to replace or modify words in the source image with another one while maintaining its realistic appearance. Specifically, we build our model on a diffusion model and carefully modify the network structure to enable the model for drawing multilingual characters with the help of glyph and position information. Moreover, we design a self-supervised learning framework to leverage large amounts of web data to improve the representation ability of the model. Experimental results show that our method achieves an impressive performance and enables controllable editing on in-the-wild images with high fidelity. Our code will be avaliable in \url{https://github.com/chenhaoxing/DiffUTE}.

Recently, numerous studies have demonstrated the effectiveness of contrastive learning (CL), which learns feature representations by pulling in positive samples while pushing away negative samples. Many successes of CL lie in that there exists semantic consistency between data augmentations of the same instance. In multi-view scenarios, however, CL might cause representation degeneration when the collected multiple views inherently have inconsistent semantic information or their representations subsequently do not capture sufficient discriminative information. To address this issue, we propose a novel framework called SEM: SElf-weighted Multi-view contrastive learning with reconstruction regularization. Specifically, SEM is a general framework where we propose to first measure the discrepancy between pairwise representations and then minimize the corresponding self-weighted contrastive loss, and thus making SEM adaptively strengthen the useful pairwise views and also weaken the unreliable pairwise views. Meanwhile, we impose a self-supervised reconstruction term to regularize the hidden features of encoders, to assist CL in accessing sufficient discriminative information of data. Experiments on public multi-view datasets verified that SEM can mitigate representation degeneration in existing CL methods and help them achieve significant performance improvements. Ablation studies also demonstrated the effectiveness of SEM with different options of weighting strategies and reconstruction terms.

Masked image modeling (MIM) has become a prevalent pre-training setup for vision foundation models and attains promising performance. Despite its success, existing MIM methods discard the decoder network during downstream applica- tions, resulting in inconsistent representations between pre-training and fine-tuning and can hamper downstream task performance. In this paper, we propose a new architecture, RevColV2, which tackles this issue by keeping the entire autoen- coder architecture during both pre-training and fine-tuning. The main body of RevColV2 contains bottom-up columns and top-down columns, between which information is reversibly propagated and gradually disentangled. Such design enables our architecture with the nice property: maintaining disentangled low-level and semantic information at the end of the network in MIM pre-training. Our experimental results suggest that a foundation model with decoupled features can achieve competitive performance across multiple downstream vision tasks such as image classification, semantic segmentation and object detection. For exam- ple, after intermediate fine-tuning on ImageNet-22K dataset, RevColV2-L attains 88.4\% top-1 accuracy on ImageNet-1K classification and 58.6 mIoU on ADE20K semantic segmentation. With extra teacher and large scale dataset, RevColv2-L achieves 62.1 APbox on COCO detection and 60.4 mIoU on ADE20K semantic segmentation.