A Comparative Evaluation of Deep Belief Nets in
Semi-supervised Learning


Abstract
In this report I studied the performance of deep belief nets (DBNs) on semi-supervised learning problems, in which only a small proportion of data are labeled. First the performance between DBNs and support vector machines (SVMs) are compared to investigate the advantage of deep models over shallow ones. I also explored the use of DBNs as pre-training for SVMs and feed-forward nets (FFNs). The experimental results show that DBN is able to yield state-of-art modeling power in semi-supervised learning.

Publication
  • Huixuan Tang, A Comparative Evaluation of Deep Belief Nets in Semi-supervised Learning, Report for CSC2515, 2008.[pdf]