Self-Contrastive Learning: Single-viewed Supervised Contrastive Framework using Sub-network
This work addresses efficiency issues in contrastive learning for computer vision researchers, offering a more computationally efficient method with competitive performance gains.
The paper tackles the high computational cost of multi-viewed supervised contrastive learning by introducing a single-viewed framework that uses a multi-exit architecture to generate multiple features from a single image, achieving a +0.6% accuracy improvement on ImageNet with ResNet-50 while reducing memory usage by 59% and training time by 48%.
Contrastive loss has significantly improved performance in supervised classification tasks by using a multi-viewed framework that leverages augmentation and label information. The augmentation enables contrast with another view of a single image but enlarges training time and memory usage. To exploit the strength of multi-views while avoiding the high computation cost, we introduce a multi-exit architecture that outputs multiple features of a single image in a single-viewed framework. To this end, we propose Self-Contrastive (SelfCon) learning, which self-contrasts within multiple outputs from the different levels of a single network. The multi-exit architecture efficiently replaces multi-augmented images and leverages various information from different layers of a network. We demonstrate that SelfCon learning improves the classification performance of the encoder network, and empirically analyze its advantages in terms of the single-view and the sub-network. Furthermore, we provide theoretical evidence of the performance increase based on the mutual information bound. For ImageNet classification on ResNet-50, SelfCon improves accuracy by +0.6% with 59% memory and 48% time of Supervised Contrastive learning, and a simple ensemble of multi-exit outputs boosts performance up to +1.5%. Our code is available at https://github.com/raymin0223/self-contrastive-learning.