Pushing the limits of self-supervised ResNets: Can we outperform supervised learning without labels on ImageNet?

Abstract

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 performancecritical settings. Building on prior theoretical insights (Mitrovic et al., 2021) we propose RELICv2 which combines an explicit invariance loss with a contrastive objective over a varied set of appropriately constructed data views. RELICv2 achieves 77.1% top-1 classification accuracy on ImageNet using linear evaluation with a ResNet50 architecture and 80.6% with larger ResNet models, outperforming previous state-ofthe-art self-supervised approaches by a wide margin. Most notably, RELICv2 is the first representation learning method to consistently outperform the supervised baseline in a like-for-like comparison using a range of standard ResNet architectures. Finally we show that despite using ResNet encoders, RELICv2 is comparable to state-of-theart self-supervised vision transformers.

The combination of Hebbian and predictive plasticity learns invariant object representations in deep sensory networks

“…Self-supervised learning (SSL), a family of unsupervised machine learning algorithms, may offer a remedy for the latter issue. Crucially, SSL does not need labeled data but instead requires that the internal network representations belonging to related inputs predict one another [15][16][17][18]. In other words, network representations themselves act as targets for similar inputs.…”

Provable Stochastic Optimization for Global Contrastive Learning: Small Batch Does Not Harm Performance

“…For a fair comparison, we adopt the same settings as SimCLR to SogCLR unless noted (main difference is the batch size). It is not our focus to leverage multiple techniques for achieving stateof-the-art performance [31]. We also compare with the CLIP framework for bimodal constrastive learning.…”

“…The performance of top-1 linear evaluation accuracy was further improved in several works [1,12,25,5,31]. For example, [1] uses clustering assignment (codes) and prediction probabilities over clusters based on learned features to define a cross-entropy loss, where the clustering assignments of each sample are computed online using the prototype vectors and the learned representations.…”

“…With a larger number of hidden nodes in MLP layers, their top-1 linear evaluation accuracy is 74.3% using a ResNet-50 backbone. Recently, [31] attains state-of-the-art performance of top-1 linear evaluation accuracy (77.1%) by using a combination of techniques, e.g., the contrastive loss as in SimCLR, the online-target setup as in [12], the multi-crop augmentations as in [1], and the invariance loss as in [25]. We also notice that larger backbones lead to better performance, e.g, ResNet-200 [12,31] and Vision Transformer [5,21].…”

“…Recently, [31] attains state-of-the-art performance of top-1 linear evaluation accuracy (77.1%) by using a combination of techniques, e.g., the contrastive loss as in SimCLR, the online-target setup as in [12], the multi-crop augmentations as in [1], and the invariance loss as in [25]. We also notice that larger backbones lead to better performance, e.g, ResNet-200 [12,31] and Vision Transformer [5,21]. We would like to point out that it is not our focus to leverage these different techniques to achieve state-ofthe-art performance.…”