Unsupervised Pre-Training with Spiking Neural Networks in Semi-Supervised Learning

Abstract

Semi-supervised learning appeared when people understood that ignoring the possibility of getting benefit from unlabeled data during supervised learning with little labeled data is not wise. It is difficult to overestimate the importance of this as much more unlabeled data exists than labeled and much easier it can be collected. Many semi-supervised learning methods were developed over the past decades. A new approach is presented in this paper, which is based on using spiking neural networks in the pre-training phase. Spiking neural networks are biologically plausible neural networks, which try to simulate the behavior of neurons and processes which occur in biological neural networks. Most of the learning rules used in spiking neural networks are unsupervised as unsupervised learning is thought to be a major drive for developmental plasticity in the brain. It is considered that it is hard for the brain to do supervised learning, things like doing math or classification. In a nutshell, the proposed method is a combination of the advantages of both spiking neuron networks (unsupervised learning) and classical artificial neural networks (supervised learning). We showed that such approach may increase the accuracy of the classifier when a small amount of labeled data is given.