Abstract

This paper describes a family of probabilistic architectures designed foronline learning under the logarithmic loss. Rather than relying on non-lineartransfer functions, our method gains representational power by the use of dataconditioning. We state under general conditions a learnable capacity theoremthat shows this approach can in principle learn any bounded Borel-measurablefunction on a compact subset of euclidean space; the result is stronger thanmany universality results for connectionist architectures because we provideboth the model and the learning procedure for which convergence is guaranteed.

 

Quick Read (beta)