From saymukh@anise.ee.cornell.edu Wed Nov 27 15:01:35 1996 Subject: Latest Workshop Program and some information Dear Workshop participants, I have enclosed the latest workshop program. Please be advised that talks are 20 minutes long, with about 5 more minutes for questions. If any questions are deemed worthy of extensive discussion, we shall deal with them in the discussion sessions. .............. ONE-DAY WORKSHOP: MODELING ERROR SURFACES Saturday, December 7, 1996 Workshop Chairs Sayandev Mukherjee* & Terrence L. Fine School of Electrical Engineering Cornell University, Ithaca, NY 14853 {saymukh,tlfine}@ee.cornell.edu INTENDED AUDIENCE (a) Those interested in algorithms for finding minima of (error) surfaces in high-dimensional spaces. (b) Those interested in effective statistical models for high-dimensional surfaces. SUMMARY The training error surface, given by the mean squared error made by the neural network when trained upon a finite training set, may have a number of minima growing exponentially in the size of the training set, while the generalization error surface of the neural network also may have multiple local minima. Clearly, a successful model for the generalization and/or training error surface would be of great help in training and model selection. However, all the conventional models for surfaces (geostatistiques models, multivariate splines, projection-pursuit models) are ineffective for the high-dimensional spaces characteristic of neural networks. Certain known properties of the topology and minima of the error surface(s) are presented and discussed in the light of their applicability to developing more tractable models. GOAL We propose a one-day workshop so as to share our knowledge about the local minima and topology of the error surface(s), and to increase our awareness of the main stumbling blocks in the development of viable models for the error surface. Our hope is that at the end of the workshop, the participants will have a better understanding of the problems involved and be ready to tackle them. PLANNED FORMAT AND TENTATIVE LIST OF SPEAKERS Morning Session (3 hours): 1. Introduction and overview: Sayandev Mukherjee 2. Insights into local minima (a) The XOR problem has no local minima: Len Hamey (b) The training error surface may have exponentially many local minima: Manfred Warmuth (c) Local minima and their effect on the convergence properties of backpropagation: Steve Lawrence 3. Group Discussion: Relation between error surface models and local minima 4. Statistical Models for the Error Surface: (a) Application of the Functional Central Limit Theorem and comparisons of local minima of Gaussian random fields: Michael Turmon Afternoon Session (3 hours): (b) Toward more convex error surfaces: Grace Wahba 5. Group Discussion: Towards effective models for high-dimensional surfaces 6. Topological, algebraic and geometric investigations of the properties of the error surface (a) Topology and geometry of single hidden layer network, least squares weight solutions: Frans Coetzee or Virginia Stonick (b) Critical points and local minima for error functions: Eduardo Sontag (c) On convexity and neural nets with a large number of hidden units: Andrew Barron 7. Panel Discussion: Future directions (Speakers+Peter Bartlett) 8. Summary and round-up: Terrence Fine ___________________________________ * Contact; email address saymukh@ee.cornell.edu