Login
Talks & Seminars
Training multilayer perceptrons
Dr. Pinaki Roy Chowdhury, Defence Research and Development Organization
Date & Time: September 24, 2002 15:30
Venue: CSE
Abstract:
The present work proposes a new framework for training multilayer perceptrons (MLP). The novelty of this scheme lie in detrapping the local minimum point computed by error backpropagation (EBP); by employing dynamic tunneling (DT) in the search space. Thus the new training methodology forms a two-phased search process that implements both gradient based and direct search alternately in the search space. The study of both learning and generalization capability of MLP trained by\NEBPDT is performed. The mathematical analysis of this framework suggests that MLP trained by EBPDT acts as an automatic regularizer, thereby producing smoothing effect in the output behaviour. This smoothing effect in turn is responsible for good generalization in unseen data. The classification results obtained on machine learning data-sets, both in crisp and fuzzy domain confirms the robust nature of the EBPDT algorithm as a learning paradigm for MLP.
Speaker Profile:
Dr. Pinaki Roy Chowdhury, Scientist Defence Terrain Research Laboratory Defence Research and Development Organization
List of Talks

Webmail

Username:
Password:
Faculty CSE IT
Forgot Password
    [+] Sitemap     Feedback