Two novel ensemble approaches for improving classification of neural networks

摘要

The task of pattern recognition is one of the most recurrent tasks that we encounter in our lives. Therefore, there has been a significant interest of automating this task for many decades. Many techniques have been developed to this end, such as neural networks. Neural networks are excellent pattern classifiers with very robust means of learning and a relatively high classification power. Naturally, there has been an increasing interest in further improving neural networks' classification for complex problems. Many methods have been proposed.;In this thesis, we propose two novel ensemble approaches to further improving neural networks' classification power, namely paralleling neural networks and chaining neural networks. The first seeks to improve a neural network's classification by combining the outputs of a set of neural networks together via another neural network. The second improves a neural network's accuracy by feeding the outputs of a neural network into another and continually doing so in a chaining fashion until the error is reduced sufficiently. The effectiveness of both approaches has been demonstrated through a series of experiments.