David Cohn to speak in Colorado Machine Learning series

[Lorien Y. Pratt]

			      The spring, 1993
			  Colorado Machine Learning 
			      Colloquium Series
			       
				 presents:

			       Dr. David Cohn
		      Dept. of Brain & Cognitive Science
		      Massachusetts Inst. of Technology
			      Cambridge, MA 02139

		Uncertainty-Based Queries in Neural Networks

			*Thursday*, April 15, 1993
		 Room 110, Stratton Hall, on the CSM campus
				  5:30 pm

				 ABSTRACT

      In many interesting learning problems, it is practical for a learner
      to pick its own training data.  Intuition and theory both suggest
      that by properly picking where one's training data comes from, one
      can greatly improve one's ability to generalize.  I will consider
      the problem of attempting to learn a map from points in a domain,
      such as a geometric map or a set of state-action pairs, to some
      ``value,'' such as a classification or a next-state identifier.
      I will assume that training data may be obtained by querying. That
      is, we may specify a point x and call an oracle, or perform an
      experiment to determine f(x), the value of the map at that point.

      We wish to make queries that are ``optimally informative'' according
      to some criterion, but there are many criteria to choose from, and
      many are computationally intractible.

      I will discuss my current research studying the suitability of a
      querying criterion based on uncertainty in system parameters.
      Under certain reasonable assumptions, one may efficiently compute
      how much the uncertainty in system parameters will be reduced by
      knowing f(x) for a specified x, and thus the ``information gain''
      of querying it. This approach was introduced by Fedorov in 1972;
      its utility for active data selection in neural networks was proposed
      by MacKay in 1991.

      I will discuss experiments performing gradient ascent on the
      information gain of a query, and discuss the problems that are
      involved in extending this approach to learning problems which only
      allow restricted querying, such as navigation, exploration and
      control.

      Suggested background readings: Training Connectionist Networks with
      Queries and Selective Sampling; Les Atlas, David Cohn and Richard
      Ladner.  Advances in Neural Information Processing 2, D. Touretzky
      ed.  Morgan Kaufmann, 1990, pp. 566-573.  Constructing Hidden Units
      using Examples and Queries; Eric B. Baum and Kevin J. Lang.  Advances
      in Neural Information Processing 3, R. Lippman, J. Moody and D.
      Touretzky, eds.  Morgan Kaufmann, 1991, pp. 904- 910. The Evidence
      Framework applied to Classification Networks; David J. C. MacKay.
      Neural Computation, Volume 4, number 5, 1992, pages 698-714.
      These readings are available on reserve at the Arthur Lakes Library
      at CSM.  Ask for the reserve package for MACS570, subject: Cohn.
      Non-students can check materials out on reserve by providing a
      driver's license.

			  Open to the Public
	  Refreshments to be served at 5:00pm, prior to talk


	For more information (including a schedule of all talks in this
	series), contact:  Dr. L. Y. Pratt, CSM Dept. of Mathematical and
	Computer Sciences, lpratt at mines.colorado.edu, (303) 273-3878.  
	The speaker may be contacted at cohn at psyche.mit.edu.

				Sponsored by:
      THE CSM DEPARTMENTS OF MATHEMATICAL AND COMPUTER SCIENCES, GEOPHYSICS,
			 DIVISION OF ENGINEERING, AND
				    CRIS
The Center for Robotics and Intelligent Systems at the Colorado School 
				  of Mines