LE3 .A278 2013

2013

Silver, Danny

Acadia University

Master of Science

Masters

Computer Science

Computer Science

Deep learning is a sub-area of machine learning which uses stacked layers of Restricted Boltzmann Machines (RBM), a type of stochastic associative arti cial neural network, to develop multi-layer generative models. In previous work, Deep Belief Networks, a supervised approach of deep learning architecture (DLA), has been used to achieve two-channel learning in practical problems. In this research, our goal is to achieve multi-channel learning with purely unsupervised DLAs. We develop an unsupervised multi-channel DLA which is composed of an RBM associative memory network and multiple stacked RBM sensory channels. It learns associative representation from two or more input channels, meant to simulate hu- man sensory modalities, such that input at one channel will correctly generate the associated response at the other channels and vice versa. In this way, the system de- velops a form of supervised classi cation model meant to simulate aspects of human associative memory. Empirical studies show that an RBM with back- tting can perform as well as a Boltzmann machine with less training time. Non-paired examples can improve over- all performance of a multi-channel DLA. An unsupervised DLA with back- tting is able to achieve multi-channel associative learning if it has su cient computational capacity. Unsupervised DLA models are shown to perform at the same level of ac- curacy as equivalent supervised back-propagation networks, and they have a better ability to di erentiate representative features from noise. The empirical results pro- vide con dence that unsupervised DLAs are able to build su ciently accurate models for multi-channel situations.

The author retains copyright in this thesis. Any substantial copying or any other actions that exceed fair dealing or other exceptions in the Copyright Act require the permission of the author.

https://scholar.acadiau.ca/islandora/object/theses:306