A SOM/MLP hybrid network that uses unlabeled data to improve classification performance

摘要

The Guelph Cluster Class (GCC) system is an approach to using unlabeled data to aid in the training of a supervised neural network. A SOM picks out natural clusters in the input data that are manipulated to correspond to various sub-classes of the desired classification space. These sub-class clusters are used to classify unlabeled data so as to provide self-labeled data. The self-labeled data is used as training data for a supervised system. Experiments have demonstrated the increased classification power for GCC especially when the initial amount of labeled data is very small.