A preliminary approach to semi-supervised learning in convolutional neural networks applying “sleep-wake” cycles

The scarcity of labeled data has limited the capacity of convolutional neural networks (CNNs) until not long ago and still represents a serious problem in a number of image processing applications. Unsupervised methods have been shown to perform well in feature extraction and clustering tasks, but further investigation on unsupervised solutions for CNNs is needed. In this work, we propose a bio-inspired methodology that applies a deep generative model to help the CNN take advantage of unlabeled data and improve its classification performance. Inspired by the human “sleep-wake cycles”, the proposed method divides the learning process into sleep and waking periods. During the waking period, both the generative model and the CNN learn from real training data simultaneously. When sleep begins, none of the networks receive real data and the generative model creates a synthetic dataset from which the CNN learns. The experimental results showed that the generative model was able to teach the CNN and improve its classification performance.