IMPROVING SMALL CONVOLUTIONAL NEURAL NETWORKS WITH SEMI-SUPERVISED LEARNING

The widespread adoption of Convolutional Neural Networks in both academia and the commercial sector have led to a rise in interest of compact solutions in constrained scenarios. Even though CNNs with a large number of layers have shown outstanding results in various tasks, these large architectures are not always well suited in some situations. The best results can be obtained when training networks with many labeled samples, which is difficult for highly specialized tasks. Using semi-supervised learning techniques, the performance of small CNNs can be boosted to make them more practical, even when the labeled set is small. The paper focuses on three semi-supervised algorithms, used in two architectures with a small number of layers and shows how their performance can be improved when training on small datasets.