BAYESIAN LEARNING VERSUS OPTIMAL LEARNING

We consider the optimal performance that may be reached in the problem of learning the symmetry-breaking direction of a cloud of P = alpha N points in a N-dimensional space. The performance is measured through the overlap R-opt between the hue symmetry-breaking direction and the learnt one. Depending on the problem, the learning curves R-opt(alpha) may present discontinuities. We show that close to these, bayesian le arning is not optimal. (C) 1998 Elsevier Science B.V. All rights reser ved.