Mandzuik BISC Seminar rescheduled: 23 June 1997

Michael Lee (leem@cs.berkeley.edu)
Fri, 20 Jun 1997 17:35:22 +0200


Dear Bisc-Group,

The BISC Seminar given by Dr. Mandziuk has been rescheduled for Monday,
the 23rd of June in room 320 Soda Hall. Attached is the copy of his
abstract and title.

Michael Lee

--
_______________________________________________________________________
Michael A. Lee
Berkeley Initiative in Soft Computing
387 Soda Hall                                      Tel: +1-510-642-9827
Computer Science Division                          Fax: +1-510-642-5775
University of California                    Email: leem@cs.berkeley.edu
Berkeley, CA 94720-1776 USA       WWW: http://www.cs.berkeley.edu/~leem
_______________________________________________________________________

Speaker: Jacek Mandziuk

Affiliation: International Computer Science Institute, Berkeley and EECS Department UC Berkeley

e-mail: mandziuk@icsi.berkeley.edu

Title: Incremental Class Learning approach and its application to Handwritten Digit Recognition

Date: June 23, 1997 Location: 320 Soda Hall, University of California, Berkeley, CA

Abstract:

Incremental Class Learning (ICL) provides a feasible framework for development of scalable learning systems. Instead of learning the whole problem at once, ICL focuses on learning subproblems incrementally, one at a time, using the results of prior learning during subsequent learning, and then combining the solutions in an appropriate manner. Therefore, the ``cathastrophic interference problem'', which occurs in sequential learning, is significantly alleviated.

The ICL approach presented in this talk can be summarized as follows. Primarily the system focuses on one category and after it learns this category, it tries to identify the compact subset of features (nodes) in the hidden layers, that are crucial for recognition of this category. The system then {\em freezes} these crucial nodes (features) by fixing their incoming weights. As a results, they cannot be obliterated in subsequent learning. These frozen features are availiable for learning other categories, and serve as parts of weight structures build subsequently to recognize other categories. The set of features gradually stabilizes and eventually, learning a new category requires less effort, and primarily consists of combining existing features in an appropriate manner. Eventually a large number of nodes are {\em shared} among various categories.

We present results of applying the ICL approach to the Handwritten Digit Recognition problem, based on the Spatio-Temporal representation of patterns.

Joint work with Lokendra Shastri.