Recently, flat-minima optimizers, which seek to find parameters in low-loss neighborhoods, have been shown to improve a neural network's generalization performance over stochastic and adaptive gradient-based optimizers. Two methods have received significant attention due to their scalability: 1. Stochastic Weight Averaging (SWA), and 2. Sharpness-Aware Minimization (SAM). However, there has been limited investigation into their properties and no systematic benchmarking of them across different domains. We fill this gap here by comparing the loss surfaces of the models trained with each method and through broad benchmarking across computer vision, natural language processing, and graph representation learning tasks. We discover several surprising findings from these results, which we hope will help researchers further improve deep learning optimizers, and practitioners identify the right optimizer for their problem.
Despite recent progress made by self-supervised methods in representation learning with residual networks, they still underperform supervised learning on the ImageNet classification benchmark, limiting their applicability in performance-critical settings. Building on prior theoretical insights from ReLIC [Mitrovic et al., 2021], we include additional inductive biases into self-supervised learning. We propose a new self-supervised representation learning method, ReLICv2, which combines an explicit invariance loss with a contrastive objective over a varied set of appropriately constructed data views to avoid learning spurious correlations and obtain more informative representations. ReLICv2 achieves $77.1\%$ top-$1$ accuracy on ImageNet under linear evaluation on a ResNet50, thus improving the previous state-of-the-art by absolute $+1.5\%$; on larger ResNet models, ReLICv2 achieves up to $80.6\%$ outperforming previous self-supervised approaches with margins up to $+2.3\%$. Most notably, ReLICv2 is the first unsupervised representation learning method to consistently outperform the supervised baseline in a like-for-like comparison over a range of ResNet architectures. Using ReLICv2, we also learn more robust and transferable representations that generalize better out-of-distribution than previous work, both on image classification and semantic segmentation. Finally, we show that despite using ResNet encoders, ReLICv2 is comparable to state-of-the-art self-supervised vision transformers.
Standard contrastive learning approaches usually require a large number of negatives for effective unsupervised learning and often exhibit slow convergence. We suspect this behavior is due to the suboptimal selection of negatives used for offering contrast to the positives. We counter this difficulty by taking inspiration from support vector machines (SVMs) to present max-margin contrastive learning (MMCL). Our approach selects negatives as the sparse support vectors obtained via a quadratic optimization problem, and contrastiveness is enforced by maximizing the decision margin. As SVM optimization can be computationally demanding, especially in an end-to-end setting, we present simplifications that alleviate the computational burden. We validate our approach on standard vision benchmark datasets, demonstrating better performance in unsupervised representation learning over state-of-the-art, while having better empirical convergence properties.
We present Masked Feature Prediction (MaskFeat) for self-supervised pre-training of video models. Our approach first randomly masks out a portion of the input sequence and then predicts the feature of the masked regions. We study five different types of features and find Histograms of Oriented Gradients (HOG), a hand-crafted feature descriptor, works particularly well in terms of both performance and efficiency. We observe that the local contrast normalization in HOG is essential for good results, which is in line with earlier work using HOG for visual recognition. Our approach can learn abundant visual knowledge and drive large-scale Transformer-based models. Without using extra model weights or supervision, MaskFeat pre-trained on unlabeled videos achieves unprecedented results of 86.7% with MViT-L on Kinetics-400, 88.3% on Kinetics-600, 80.4% on Kinetics-700, 39.8 mAP on AVA, and 75.0% on SSv2. MaskFeat further generalizes to image input, which can be interpreted as a video with a single frame and obtains competitive results on ImageNet.
Contrastive representation learning has proven to be an effective self-supervised learning method. Most successful approaches are based on Noise Contrastive Estimation (NCE) and use different views of an instance as positives that should be contrasted with other instances, called negatives, that are considered as noise. However, several instances in a dataset are drawn from the same distribution and share underlying semantic information. A good data representation should contain relations, or semantic similarity, between the instances. Contrastive learning implicitly learns relations but considering all negatives as noise harms the quality of the learned relations. To circumvent this issue, we propose a novel formulation of contrastive learning using semantic similarity between instances called Similarity Contrastive Estimation (SCE). Our training objective is a soft contrastive learning one. Instead of hard classifying positives and negatives, we estimate from one view of a batch a continuous distribution to push or pull instances based on their semantic similarities. This target similarity distribution is sharpened to eliminate noisy relations. The model predicts for each instance, from another view, the target distribution while contrasting its positive with negatives. Experimental results show that SCE is Top-1 on the ImageNet linear evaluation protocol at 100 pretraining epochs with 72.1% accuracy and is competitive with state-of-the-art algorithms by reaching 75.4% for 200 epochs with multi-crop. We also show that SCE is able to generalize to several tasks. Source code is available here: https://github.com/CEA-LIST/SCE.
This paper explores a better prediction target for BERT pre-training of vision transformers. We observe that current prediction targets disagree with human perception judgment.This contradiction motivates us to learn a perceptual prediction target. We argue that perceptually similar images should stay close to each other in the prediction target space. We surprisingly find one simple yet effective idea: enforcing perceptual similarity during the dVAE training. Moreover, we adopt a self-supervised transformer model for deep feature extraction and show that it works well for calculating perceptual similarity.We demonstrate that such learned visual tokens indeed exhibit better semantic meanings, and help pre-training achieve superior transfer performance in various downstream tasks. For example, we achieve $\textbf{84.5\%}$ Top-1 accuracy on ImageNet-1K with ViT-B backbone, outperforming the competitive method BEiT by $\textbf{+1.3\%}$ under the same pre-training epochs. Our approach also gets significant improvement on object detection and segmentation on COCO and semantic segmentation on ADE20K. Equipped with a larger backbone ViT-H, we achieve the state-of-the-art ImageNet accuracy (\textbf{88.3\%}) among methods using only ImageNet-1K data.
This paper presents SimMIM, a simple framework for masked image modeling. We simplify recently proposed related approaches without special designs such as block-wise masking and tokenization via discrete VAE or clustering. To study what let the masked image modeling task learn good representations, we systematically study the major components in our framework, and find that simple designs of each component have revealed very strong representation learning performance: 1) random masking of the input image with a moderately large masked patch size (e.g., 32) makes a strong pre-text task; 2) predicting raw pixels of RGB values by direct regression performs no worse than the patch classification approaches with complex designs; 3) the prediction head can be as light as a linear layer, with no worse performance than heavier ones. Using ViT-B, our approach achieves 83.8% top-1 fine-tuning accuracy on ImageNet-1K by pre-training also on this dataset, surpassing previous best approach by +0.6%. When applied on a larger model of about 650 million parameters, SwinV2-H, it achieves 87.1% top-1 accuracy on ImageNet-1K using only ImageNet-1K data. We also leverage this approach to facilitate the training of a 3B model (SwinV2-G), that by $40\times$ less data than that in previous practice, we achieve the state-of-the-art on four representative vision benchmarks. The code and models will be publicly available at https://github.com/microsoft/SimMIM.
The success of language Transformers is primarily attributed to the pretext task of masked language modeling (MLM), where texts are first tokenized into semantically meaningful pieces. In this work, we study masked image modeling (MIM) and indicate the advantages and challenges of using a semantically meaningful visual tokenizer. We present a self-supervised framework iBOT that can perform masked prediction with an online tokenizer. Specifically, we perform self-distillation on masked patch tokens and take the teacher network as the online tokenizer, along with self-distillation on the class token to acquire visual semantics. The online tokenizer is jointly learnable with the MIM objective and dispenses with a multi-stage training pipeline where the tokenizer needs to be pre-trained beforehand. We show the prominence of iBOT by achieving an 82.3% linear probing accuracy and an 87.8% fine-tuning accuracy evaluated on ImageNet-1K. Beyond the state-of-the-art image classification results, we underline emerging local semantic patterns, which helps the models to obtain strong robustness against common corruptions and achieve leading results on dense downstream tasks, eg., object detection, instance segmentation, and semantic segmentation.
This paper shows that masked autoencoders (MAE) are scalable self-supervised learners for computer vision. Our MAE approach is simple: we mask random patches of the input image and reconstruct the missing pixels. It is based on two core designs. First, we develop an asymmetric encoder-decoder architecture, with an encoder that operates only on the visible subset of patches (without mask tokens), along with a lightweight decoder that reconstructs the original image from the latent representation and mask tokens. Second, we find that masking a high proportion of the input image, e.g., 75%, yields a nontrivial and meaningful self-supervisory task. Coupling these two designs enables us to train large models efficiently and effectively: we accelerate training (by 3x or more) and improve accuracy. Our scalable approach allows for learning high-capacity models that generalize well: e.g., a vanilla ViT-Huge model achieves the best accuracy (87.8%) among methods that use only ImageNet-1K data. Transfer performance in downstream tasks outperforms supervised pre-training and shows promising scaling behavior.
We present TWIST, a simple and theoretically explainable self-supervised representation learning method by classifying large-scale unlabeled datasets in an end-to-end way. We employ a siamese network terminated by a softmax operation to produce twin class distributions of two augmented images. Without supervision, we enforce the class distributions of different augmentations to be consistent. However, simply minimizing the divergence between augmentations will cause collapsed solutions, i.e., outputting the same class probability distribution for all images. In this case, no information about the input image is left. To solve this problem, we propose to maximize the mutual information between the input and the class predictions. Specifically, we minimize the entropy of the distribution for each sample to make the class prediction for each sample assertive and maximize the entropy of the mean distribution to make the predictions of different samples diverse. In this way, TWIST can naturally avoid the collapsed solutions without specific designs such as asymmetric network, stop-gradient operation, or momentum encoder. As a result, TWIST outperforms state-of-the-art methods on a wide range of tasks. Especially, TWIST performs surprisingly well on semi-supervised learning, achieving 61.2% top-1 accuracy with 1% ImageNet labels using a ResNet-50 as backbone, surpassing previous best results by an absolute improvement of 6.2%. Codes and pre-trained models are given on: https://github.com/bytedance/TWIST
Unsupervised visual representation learning has gained much attention from the computer vision community because of the recent achievement of contrastive learning. Most of the existing contrastive learning frameworks adopt the instance discrimination as the pretext task, which treating every single instance as a different class. However, such method will inevitably cause class collision problems, which hurts the quality of the learned representation. Motivated by this observation, we introduced a weakly supervised contrastive learning framework (WCL) to tackle this issue. Specifically, our proposed framework is based on two projection heads, one of which will perform the regular instance discrimination task. The other head will use a graph-based method to explore similar samples and generate a weak label, then perform a supervised contrastive learning task based on the weak label to pull the similar images closer. We further introduced a K-Nearest Neighbor based multi-crop strategy to expand the number of positive samples. Extensive experimental results demonstrate WCL improves the quality of self-supervised representations across different datasets. Notably, we get a new state-of-the-art result for semi-supervised learning. With only 1\% and 10\% labeled examples, WCL achieves 65\% and 72\% ImageNet Top-1 Accuracy using ResNet50, which is even higher than SimCLRv2 with ResNet101.
Learning effective visual representations that generalize well without human supervision is a fundamental problem in order to apply Machine Learning to a wide variety of tasks. Recently, two families of self-supervised methods, contrastive learning and latent bootstrapping, exemplified by SimCLR and BYOL respectively, have made significant progress. In this work, we hypothesize that adding explicit information compression to these algorithms yields better and more robust representations. We verify this by developing SimCLR and BYOL formulations compatible with the Conditional Entropy Bottleneck (CEB) objective, allowing us to both measure and control the amount of compression in the learned representation, and observe their impact on downstream tasks. Furthermore, we explore the relationship between Lipschitz continuity and compression, showing a tractable lower bound on the Lipschitz constant of the encoders we learn. As Lipschitz continuity is closely related to robustness, this provides a new explanation for why compressed models are more robust. Our experiments confirm that adding compression to SimCLR and BYOL significantly improves linear evaluation accuracies and model robustness across a wide range of domain shifts. In particular, the compressed version of BYOL achieves 76.0% Top-1 linear evaluation accuracy on ImageNet with ResNet-50, and 78.8% with ResNet-50 2x.
Self-supervised Learning (SSL) including the mainstream contrastive learning has achieved great success in learning visual representations without data annotations. However, most of methods mainly focus on the instance level information (\ie, the different augmented images of the same instance should have the same feature or cluster into the same class), but there is a lack of attention on the relationships between different instances. In this paper, we introduced a novel SSL paradigm, which we term as relational self-supervised learning (ReSSL) framework that learns representations by modeling the relationship between different instances. Specifically, our proposed method employs sharpened distribution of pairwise similarities among different instances as \textit{relation} metric, which is thus utilized to match the feature embeddings of different augmentations. Moreover, to boost the performance, we argue that weak augmentations matter to represent a more reliable relation, and leverage momentum strategy for practical efficiency. Experimental results show that our proposed ReSSL significantly outperforms the previous state-of-the-art algorithms in terms of both performance and training efficiency. Code is available at \url{https://github.com/KyleZheng1997/ReSSL}.
This paper investigates two techniques for developing efficient self-supervised vision transformers (EsViT) for visual representation learning. First, we show through a comprehensive empirical study that multi-stage architectures with sparse self-attentions can significantly reduce modeling complexity but with a cost of losing the ability to capture fine-grained correspondences between image regions. Second, we propose a new pre-training task of region matching which allows the model to capture fine-grained region dependencies and as a result significantly improves the quality of the learned vision representations. Our results show that combining the two techniques, EsViT achieves 81.3% top-1 on the ImageNet linear probe evaluation, outperforming prior arts with around an order magnitude of higher throughput. When transferring to downstream linear classification tasks, EsViT outperforms its supervised counterpart on 17 out of 18 datasets. The code and models are publicly available: https://github.com/microsoft/esvit
Following their success in natural language processing, transformers have recently shown much promise for computer vision. The self-attention operation underlying transformers yields global interactions between all tokens ,i.e. words or image patches, and enables flexible modelling of image data beyond the local interactions of convolutions. This flexibility, however, comes with a quadratic complexity in time and memory, hindering application to long sequences and high-resolution images. We propose a "transposed" version of self-attention that operates across feature channels rather than tokens, where the interactions are based on the cross-covariance matrix between keys and queries. The resulting cross-covariance attention (XCA) has linear complexity in the number of tokens, and allows efficient processing of high-resolution images. Our cross-covariance image transformer (XCiT) is built upon XCA. It combines the accuracy of conventional transformers with the scalability of convolutional architectures. We validate the effectiveness and generality of XCiT by reporting excellent results on multiple vision benchmarks, including image classification and self-supervised feature learning on ImageNet-1k, object detection and instance segmentation on COCO, and semantic segmentation on ADE20k.
We introduce a self-supervised vision representation model BEiT, which stands for Bidirectional Encoder representation from Image Transformers. Following BERT developed in the natural language processing area, we propose a masked image modeling task to pretrain vision Transformers. Specifically, each image has two views in our pre-training, i.e, image patches (such as 16x16 pixels), and visual tokens (i.e., discrete tokens). We first "tokenize" the original image into visual tokens. Then we randomly mask some image patches and fed them into the backbone Transformer. The pre-training objective is to recover the original visual tokens based on the corrupted image patches. After pre-training BEiT, we directly fine-tune the model parameters on downstream tasks by appending task layers upon the pretrained encoder. Experimental results on image classification and semantic segmentation show that our model achieves competitive results with previous pre-training methods. For example, base-size BEiT achieves 83.2% top-1 accuracy on ImageNet-1K, significantly outperforming from-scratch DeiT training (81.8%) with the same setup. Moreover, large-size BEiT obtains 86.3% only using ImageNet-1K, even outperforming ViT-L with supervised pre-training on ImageNet-22K (85.2%). The code and pretrained models are available at https://aka.ms/beit.
Cluster discrimination is an effective pretext task for unsupervised representation learning, which often consists of two phases: clustering and discrimination. Clustering is to assign each instance a pseudo label that will be used to learn representations in discrimination. The main challenge resides in clustering since prevalent clustering methods (e.g., k-means) have to run in a batch mode. Besides, there can be a trivial solution consisting of a dominating cluster. To address these challenges, we first investigate the objective of clustering-based representation learning. Based on this, we propose a novel clustering-based pretext task with online \textbf{Co}nstrained \textbf{K}-m\textbf{e}ans (\textbf{CoKe}). Compared with the balanced clustering that each cluster has exactly the same size, we only constrain the minimal size of each cluster to flexibly capture the inherent data structure. More importantly, our online assignment method has a theoretical guarantee to approach the global optimum. By decoupling clustering and discrimination, CoKe can achieve competitive performance when optimizing with only a single view from each instance. Extensive experiments on ImageNet and other benchmark data sets verify both the efficacy and efficiency of our proposal. Code is available at \url{https://github.com/idstcv/CoKe}.
Self-supervised learning holds promise in leveraging large amounts of unlabeled data, however much of its progress has thus far been limited to highly curated pre-training data such as ImageNet. We explore the effects of contrastive learning from larger, less-curated image datasets such as YFCC, and find there is indeed a large difference in the resulting representation quality. We hypothesize that this curation gap is due to a shift in the distribution of image classes -- which is more diverse and heavy-tailed -- resulting in less relevant negative samples to learn from. We test this hypothesis with a new approach, Divide and Contrast (DnC), which alternates between contrastive learning and clustering-based hard negative mining. When pretrained on less curated datasets, DnC greatly improves the performance of self-supervised learning on downstream tasks, while remaining competitive with the current state-of-the-art on curated datasets.
Existing public person Re-Identification~(ReID) datasets are small in modern terms because of labeling difficulty. Although unlabeled surveillance video is abundant and relatively easy to obtain, it is unclear how to leverage these footage to learn meaningful ReID representations. In particular, most existing unsupervised and domain adaptation ReID methods utilize only the public datasets in their experiments, with labels removed. In addition, due to small data sizes, these methods usually rely on fine tuning by the unlabeled training data in the testing domain to achieve good performance. Inspired by the recent progress of large-scale self-supervised image classification using contrastive learning, we propose to learn ReID representation from large-scale unlabeled surveillance video alone. Assisted by off-the-shelf pedestrian detection tools, we apply the contrastive loss at both the image and the tracklet levels. Together with a principal component analysis step using camera labels freely available, our evaluation using a large-scale unlabeled dataset shows far superior performance among unsupervised methods that do not use any training data in the testing domain. Furthermore, the accuracy improves with the data size and therefore our method has great potential with even larger and more diversified datasets.
Recent self-supervised methods for image representation learning are based on maximizing the agreement between embedding vectors from different views of the same image. A trivial solution is obtained when the encoder outputs constant vectors. This collapse problem is often avoided through implicit biases in the learning architecture, that often lack a clear justification or interpretation. In this paper, we introduce VICReg (Variance-Invariance-Covariance Regularization), a method that explicitly avoids the collapse problem with a simple regularization term on the variance of the embeddings along each dimension individually. VICReg combines the variance term with a decorrelation mechanism based on redundancy reduction and covariance regularization, and achieves results on par with the state of the art on several downstream tasks. In addition, we show that incorporating our new variance term into other methods helps stabilize the training and leads to performance improvements.