CNN-JEPA: Self-Supervised Pretraining Convolutional Neural Networks Using Joint Embedding Predictive Architecture

TL;DR

CNN-JEPA introduces a self-supervised learning method for CNNs that improves efficiency and accuracy, outperforming previous CNN SSL methods and approaching the performance of transformer-based models with less training time.

Contribution

The paper presents CNN-JEPA, a novel SSL approach tailored for CNNs that simplifies training, reduces time, and enhances performance compared to existing CNN SSL methods.

Findings

CNN-JEPA achieves 73.3% linear top-1 accuracy on ImageNet-100 with ResNet-50.

It requires 17-35% less training time than comparable CNN SSL methods.

CNN-JEPA approaches the accuracy of methods like BYOL, SimCLR, and VICReg.

Abstract

Self-supervised learning (SSL) has become an important approach in pretraining large neural networks, enabling unprecedented scaling of model and dataset sizes. While recent advances like I-JEPA have shown promising results for Vision Transformers, adapting such methods to Convolutional Neural Networks (CNNs) presents unique challenges. In this paper, we introduce CNN-JEPA, a novel SSL method that successfully applies the joint embedding predictive architecture approach to CNNs. Our method incorporates a sparse CNN encoder to handle masked inputs, a fully convolutional predictor using depthwise separable convolutions, and an improved masking strategy. We demonstrate that CNN-JEPA outperforms I-JEPA with ViT architectures on ImageNet-100, achieving a 73.3% linear top-1 accuracy using a standard ResNet-50 encoder. Compared to other CNN-based SSL methods, CNN-JEPA requires 17-35% less training time for the same number of epochs and approaches the linear and k-NN top-1 accuracies of BYOL, SimCLR, and VICReg. Our approach offers a simpler, more efficient alternative to existing SSL methods for CNNs, requiring minimal augmentations and no separate projector network.