Self-supervised learning (SSL) has achieved promising downstream performance. However, when facing various resource budgets in real-world applications, it costs a huge computation burden to pretrain multiple networks of various sizes one by one. In this paper, we propose Discriminative-SSL-based Slimmable Pretrained Networks (DSPNet), which can be trained at once and then slimmed to multiple sub-networks of various sizes, each of which faithfully learns good representation and can serve as good initialization for downstream tasks with various resource budgets. Specifically, we extend the idea of slimmable networks to a discriminative SSL paradigm, by integrating SSL and knowledge distillation gracefully. We show comparable or improved performance of DSPNet on ImageNet to the networks individually pretrained one by one under the linear evaluation and semi-supervised evaluation protocols, while reducing large training cost. The pretrained models also generalize well on downstream detection and segmentation tasks. Code will be made public.
By leveraging contrastive learning, clustering, and other pretext tasks, unsupervised methods for learning image representations have reached impressive results on standard benchmarks. The result has been a crowded field - many methods with substantially different implementations yield results that seem nearly identical on popular benchmarks, such as linear evaluation on ImageNet. However, a single result does not tell the whole story. In this paper, we compare methods using performance-based benchmarks such as linear evaluation, nearest neighbor classification, and clustering for several different datasets, demonstrating the lack of a clear front-runner within the current state-of-the-art. In contrast to prior work that performs only supervised vs. unsupervised comparison, we compare several different unsupervised methods against each other. To enrich this comparison, we analyze embeddings with measurements such as uniformity, tolerance, and centered kernel alignment (CKA), and propose two new metrics of our own: nearest neighbor graph similarity and linear prediction overlap. We reveal through our analysis that in isolation, single popular methods should not be treated as though they represent the field as a whole, and that future work ought to consider how to leverage the complimentary nature of these methods. We also leverage CKA to provide a framework to robustly quantify augmentation invariance, and provide a reminder that certain types of invariance will be undesirable for downstream tasks.
We present federated momentum contrastive clustering (FedMCC), a learning framework that can not only extract discriminative representations over distributed local data but also perform data clustering. In FedMCC, a transformed data pair passes through both the online and target networks, resulting in four representations over which the losses are determined. The resulting high-quality representations generated by FedMCC can outperform several existing self-supervised learning methods for linear evaluation and semi-supervised learning tasks. FedMCC can easily be adapted to ordinary centralized clustering through what we call momentum contrastive clustering (MCC). We show that MCC achieves state-of-the-art clustering accuracy results in certain datasets such as STL-10 and ImageNet-10. We also present a method to reduce the memory footprint of our clustering schemes.
Self-Supervised learning (SSL) has become the new state-of-art in several domain classification and segmentation tasks. Of these, one popular category in SSL is distillation networks such as BYOL. This work proposes RSDnet, which applies the distillation network (BYOL) in the remote sensing (RS) domain where data is non-trivially different from natural RGB images. Since Multi-spectral (MS) and synthetic aperture radar (SAR) sensors provide varied spectral and spatial resolution information, we utilised them as an implicit augmentation to learn invariant feature embeddings. In order to learn RS based invariant features with SSL, we trained RSDnet in two ways, i.e., single channel feature learning and three channel feature learning. This work explores the usefulness of single channel feature learning from random MS and SAR bands compared to the common notion of using three or more bands. In our linear evaluation, these single channel features reached a 0.92 F1 score on the EuroSAT classification task and 59.6 mIoU on the DFC segmentation task for certain single bands. We also compared our results with ImageNet weights and showed that the RS based SSL model outperforms the supervised ImageNet based model. We further explored the usefulness of multi-modal data compared to single modality data, and it is shown that utilising MS and SAR data learn better invariant representations than utilising only MS data.
Self-supervised learning (SSL) is capable of learning remarkable representations from centrally available data. Recent works further implement federated learning with SSL to learn from rapidly growing decentralized unlabeled images (e.g., from cameras and phones), often resulted from privacy constraints. Extensive attention has been paid to SSL approaches based on Siamese networks. However, such an effort has not yet revealed deep insights into various fundamental building blocks for the federated self-supervised learning (FedSSL) architecture. We aim to fill in this gap via in-depth empirical study and propose a new method to tackle the non-independently and identically distributed (non-IID) data problem of decentralized data. Firstly, we introduce a generalized FedSSL framework that embraces existing SSL methods based on Siamese networks and presents flexibility catering to future methods. In this framework, a server coordinates multiple clients to conduct SSL training and periodically updates local models of clients with the aggregated global model. Using the framework, our study uncovers unique insights of FedSSL: 1) stop-gradient operation, previously reported to be essential, is not always necessary in FedSSL; 2) retaining local knowledge of clients in FedSSL is particularly beneficial for non-IID data. Inspired by the insights, we then propose a new approach for model update, Federated Divergence-aware Exponential Moving Average update (FedEMA). FedEMA updates local models of clients adaptively using EMA of the global model, where the decay rate is dynamically measured by model divergence. Extensive experiments demonstrate that FedEMA outperforms existing methods by 3-4% on linear evaluation. We hope that this work will provide useful insights for future research.
One of the latest self-supervised learning (SSL) methods, VICReg, showed a great performance both in the linear evaluation and the fine-tuning evaluation. However, VICReg is proposed in computer vision and it learns by pulling representations of random crops of an image while maintaining the representation space by the variance and covariance loss. However, VICReg would be ineffective on non-stationary time series where different parts/crops of input should be differently encoded to consider the non-stationarity. Another recent SSL proposal, Temporal Neighborhood Coding (TNC) is effective for encoding non-stationary time series. This study shows that a combination of a VICReg-style method and TNC is very effective for SSL on non-stationary time series, where a non-stationary seismic signal time series is used as an evaluation dataset.
Self-supervised learning (SSL) has shown impressive results in downstream classification tasks. However, there is limited work in understanding their failure modes and interpreting their learned representations. In this paper, we study the representation space of state-of-the-art self-supervised models including SimCLR, SwaV, MoCo, BYOL, DINO, SimSiam, VICReg and Barlow Twins. Without the use of class label information, we discover discriminative features that correspond to unique physical attributes in images, present mostly in correctly-classified representations. Using these features, we can compress the representation space by up to 40% without significantly affecting linear classification performance. We then propose Self-Supervised Representation Quality Score (or Q-Score), an unsupervised score that can reliably predict if a given sample is likely to be mis-classified during linear evaluation, achieving AUPRC of 91.45 on ImageNet-100 and 78.78 on ImageNet-1K. Q-Score can also be used as a regularization term on pre-trained encoders to remedy low-quality representations. Fine-tuning with Q-Score regularization can boost the linear probing accuracy of SSL models by up to 5.8% on ImageNet-100 and 3.7% on ImageNet-1K compared to their baselines. Finally, using gradient heatmaps and Salient ImageNet masks, we define a metric to quantify the interpretability of each representation. We show that discriminative features are strongly correlated to core attributes and, enhancing these features through Q-score regularization makes SSL representations more interpretable.
We propose Fuse Local and Global Semantics in Representation Learning (FLAGS) to generate richer representations. FLAGS aims at extract both global and local semantics from images to benefit various downstream tasks. It shows promising results under common linear evaluation protocol. We also conduct detection and segmentation on PASCAL VOC and COCO to show the representations extracted by FLAGS are transferable.
In this work, we study self-supervised representation learning for 3D skeleton-based action recognition. We extend Bootstrap Your Own Latent (BYOL) for representation learning on skeleton sequence data and propose a new data augmentation strategy including two asymmetric transformation pipelines. We also introduce a multi-viewpoint sampling method that leverages multiple viewing angles of the same action captured by different cameras. In the semi-supervised setting, we show that the performance can be further improved by knowledge distillation from wider networks, leveraging once more the unlabeled samples. We conduct extensive experiments on the NTU-60 and NTU-120 datasets to demonstrate the performance of our proposed method. Our method consistently outperforms the current state of the art on both linear evaluation and semi-supervised benchmarks.
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.
Pre-training a model to learn transferable video-text representation for retrieval has attracted a lot of attention in recent years. Previous dominant works mainly adopt two separate encoders for efficient retrieval, but ignore local associations between videos and texts. Another line of research uses a joint encoder to interact video with texts, but results in low efficiency since each text-video pair needs to be fed into the model. In this work, we enable fine-grained video-text interactions while maintaining high efficiency for retrieval via a novel pretext task, dubbed as Multiple Choice Questions (MCQ), where a parametric module BridgeFormer is trained to answer the "questions" constructed by the text features via resorting to the video features. Specifically, we exploit the rich semantics of text (i.e., nouns and verbs) to build questions, with which the video encoder can be trained to capture more regional content and temporal dynamics. In the form of questions and answers, the semantic associations between local video-text features can be properly established. BridgeFormer is able to be removed for downstream retrieval, rendering an efficient and flexible model with only two encoders. Our method outperforms state-of-the-art methods on the popular text-to-video retrieval task in five datasets with different experimental setups (i.e., zero-shot and fine-tune), including HowTo100M (one million videos). We further conduct zero-shot action recognition, which can be cast as video-to-text retrieval, and our approach also significantly surpasses its counterparts. As an additional benefit, our method achieves competitive results with much shorter pre-training videos on single-modality downstream tasks, e.g., action recognition with linear evaluation.
We introduce Consistent Assignment for Representation Learning (CARL), an unsupervised learning method to learn visual representations by combining ideas from self-supervised contrastive learning and deep clustering. By viewing contrastive learning from a clustering perspective, CARL learns unsupervised representations by learning a set of general prototypes that serve as energy anchors to enforce different views of a given image to be assigned to the same prototype. Unlike contemporary work on contrastive learning with deep clustering, CARL proposes to learn the set of general prototypes in an online fashion, using gradient descent without the necessity of using non-differentiable algorithms or K-Means to solve the cluster assignment problem. CARL surpasses its competitors in many representations learning benchmarks, including linear evaluation, semi-supervised learning, and transfer learning.
Self-supervised learning has shown its great potential to extract powerful visual representations without human annotations. Various works are proposed to deal with self-supervised learning from different perspectives: (1) contrastive learning methods (e.g., MoCo, SimCLR) utilize both positive and negative samples to guide the training direction; (2) asymmetric network methods (e.g., BYOL, SimSiam) get rid of negative samples via the introduction of a predictor network and the stop-gradient operation; (3) feature decorrelation methods (e.g., Barlow Twins, VICReg) instead aim to reduce the redundancy between feature dimensions. These methods appear to be quite different in the designed loss functions from various motivations. The final accuracy numbers also vary, where different networks and tricks are utilized in different works. In this work, we demonstrate that these methods can be unified into the same form. Instead of comparing their loss functions, we derive a unified formula through gradient analysis. Furthermore, we conduct fair and detailed experiments to compare their performances. It turns out that there is little gap between these methods, and the use of momentum encoder is the key factor to boost performance. From this unified framework, we propose UniGrad, a simple but effective gradient form for self-supervised learning. It does not require a memory bank or a predictor network, but can still achieve state-of-the-art performance and easily adopt other training strategies. Extensive experiments on linear evaluation and many downstream tasks also show its effectiveness. Code is released at https://github.com/fundamentalvision/UniGrad.
While self-supervised representation learning (SSL) has proved to be effective in the large model, there is still a huge gap between the SSL and supervised method in the lightweight model when following the same solution. We delve into this problem and find that the lightweight model is prone to collapse in semantic space when simply performing instance-wise contrast. To address this issue, we propose a relation-wise contrastive paradigm with Relation Knowledge Distillation (ReKD). We introduce a heterogeneous teacher to explicitly mine the semantic information and transferring a novel relation knowledge to the student (lightweight model). The theoretical analysis supports our main concern about instance-wise contrast and verify the effectiveness of our relation-wise contrastive learning. Extensive experimental results also demonstrate that our method achieves significant improvements on multiple lightweight models. Particularly, the linear evaluation on AlexNet obviously improves the current state-of-art from 44.7% to 50.1%, which is the first work to get close to the supervised 50.5%. Code will be made available.
The pretrain-finetune paradigm is a classical pipeline in visual learning. Recent progress on unsupervised pretraining methods shows superior transfer performance to their supervised counterparts. This paper revisits this phenomenon and sheds new light on understanding the transferability gap between unsupervised and supervised pretraining from a multilayer perceptron (MLP) perspective. While previous works focus on the effectiveness of MLP on unsupervised image classification where pretraining and evaluation are conducted on the same dataset, we reveal that the MLP projector is also the key factor to better transferability of unsupervised pretraining methods than supervised pretraining methods. Based on this observation, we attempt to close the transferability gap between supervised and unsupervised pretraining by adding an MLP projector before the classifier in supervised pretraining. Our analysis indicates that the MLP projector can help retain intra-class variation of visual features, decrease the feature distribution distance between pretraining and evaluation datasets, and reduce feature redundancy. Extensive experiments on public benchmarks demonstrate that the added MLP projector significantly boosts the transferability of supervised pretraining, e.g. +7.2% top-1 accuracy on the concept generalization task, +5.8% top-1 accuracy for linear evaluation on 12-domain classification tasks, and +0.8% AP on COCO object detection task, making supervised pretraining comparable or even better than unsupervised pretraining.
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.
State-of-the-art frameworks in self-supervised learning have recently shown that fully utilizing transformer-based models can lead to performance boost compared to conventional CNN models. Striving to maximize the mutual information of two views of an image, existing works apply a contrastive loss to the final representations. Motivated by self-distillation in the supervised regime, we further exploit this by allowing the intermediate representations to learn from the final layer via the contrastive loss. Through self-distillation, the intermediate layers are better suited for instance discrimination, making the performance of an early-exited sub-network not much degraded from that of the full network. This renders the pretext task easier also for the final layer, leading to better representations. Our method, Self-Distilled Self-Supervised Learning (SDSSL), outperforms competitive baselines (SimCLR, BYOL and MoCo v3) using ViT on various tasks and datasets. In the linear evaluation and k-NN protocol, SDSSL not only leads to superior performance in the final layers, but also in most of the lower layers. Furthermore, qualitative and quantitative analyses show how representations are formed more effectively along the transformer layers. Code is available at https://github.com/hagiss/SDSSL.
Self-supervised contrastive learning frameworks have progressed rapidly over the last few years. In this paper, we propose a novel loss function for contrastive learning. We model our pre-training task as a binary classification problem to induce an implicit contrastive effect. We further improve the n\"aive loss function after removing the effect of the positive-positive repulsion and incorporating the upper bound of the negative pair repulsion. Unlike existing methods, the proposed loss function optimizes the mutual information in positive and negative pairs. We also present a closed-form expression for the parameter gradient flow and compare the behaviour of self-supervised contrastive frameworks using Hessian eigenspectrum to analytically study their convergence. The proposed method outperforms SOTA self-supervised contrastive frameworks on benchmark datasets such as CIFAR-10, CIFAR-100, STL-10, and Tiny-ImageNet. After 200 pretraining epochs with ResNet-18 as the backbone, the proposed model achieves an accuracy of 86.36%, 58.18%, 80.50%, and 30.87% on the CIFAR-10, CIFAR-100, STL-10, and Tiny-ImageNet datasets, respectively, and surpasses the SOTA contrastive baseline by 1.93%, 3.57%, 4.85%, and 0.33%, respectively. The proposed framework also achieves a state-of-the-art accuracy of 78.4% (200 epochs) and 65.22% (100 epochs) Top-1 Linear Evaluation accuracy on ImageNet100 and ImageNet1K datasets, respectively.
Existing Class Incremental Learning (CIL) methods are based on a supervised classification framework sensitive to data labels. When updating them based on the new class data, they suffer from catastrophic forgetting: the model cannot discern old class data clearly from the new. In this paper, we explore the performance of Self-Supervised representation learning in Class Incremental Learning (SSCIL) for the first time, which discards data labels and the model's classifiers. To comprehensively discuss the difference in performance between supervised and self-supervised methods in CIL, we set up three different class incremental schemes: Random Class Scheme, Semantic Class Scheme, and Cluster Scheme, to simulate various class incremental learning scenarios. Besides, we propose Linear Evaluation Protocol (LEP) and Generalization Evaluation Protocol (GEP) to metric the model's representation classification ability and generalization in CIL. Our experiments (on ImageNet-100 and ImageNet) show that SSCIL has better anti-forgetting ability and robustness than supervised strategies in CIL. To understand what alleviates the catastrophic forgetting in SSCIL, we study the major components of SSCIL and conclude that (1) the composition of different data augmentation improves the quality of the model's representation and the \textit{Grayscale} operation reduces the system noise of data augmentation in SSCIL. (2) the projector, like a buffer, reduces unnecessary parameter updates of the model in SSCIL and increases the robustness of the model. Although the performance of SSCIL is significantly higher than supervised methods in CIL, there is still an apparent gap with joint learning. Our exploration gives a baseline of self-supervised class incremental learning on large-scale datasets and contributes some forward strategies for mitigating the catastrophic forgetting in CIL.
Contrastive Learning has shown impressive results on natural and medical images, without requiring annotated data. However, a particularity of medical images is the availability of meta-data (such as age or sex) that can be exploited for learning representations. Here, we show that the recently proposed contrastive y-Aware InfoNCE loss, that integrates multi-dimensional meta-data, asymptotically optimizes two properties: conditional alignment and global uniformity. Similarly to [Wang, 2020], conditional alignment means that similar samples should have similar features, but conditionally on the meta-data. Instead, global uniformity means that the (normalized) features should be uniformly distributed on the unit hyper-sphere, independently of the meta-data. Here, we propose to define conditional uniformity, relying on the meta-data, that repel only samples with dissimilar meta-data. We show that direct optimization of both conditional alignment and uniformity improves the representations, in terms of linear evaluation, on both CIFAR-100 and a brain MRI dataset.