A New Self-supervised Method for Supervised Learning

In traditional self-supervised visual feature learning, convolutional neural networks (ConvNets) trained by a proposed pretext task with only unlabeled data encode high-level semantic visual representations for downstream tasks of interest. The proposed pretext tasks are mostly based on images or videos. In this work, starting from the feature layers, we propose a completely new pretext task formulated within ConvNets, and use it to enhance the supervised learning of fully labeled datasets. We discard the channels on feature maps after particular convolutional layers to generate self-supervised labels, and combine them with the original labels for classification. Our objective is to mine richer feature information by making ConvNets understand which channels are missing at the same time of classification. Experiments show that our improvement is effective across multiple models and datasets.