Abstract
We present experiments on pattern classification with synchronous Boltzmann machines. These experiments allowed us to validate on a realistic problem the usefulness of this such stochastic neural networks. Experiments were run on a massively parallel computer (Connection Machine). The construction of the training set as well as the coding of the input modules were carefully designed, in order to achieve strong geometric invariance properties in the final pattern classification. Our network is composed of interacting modules, with preconstrained connections, leading to a sparse weight matrix to be learned. Experiments using modified learning rules implementing regularity constraints on this weight matrix were successfully performed. In addition, we devised methods for the interpretation of the final network weight structure, using linear regression, discriminant and principal component analyses.This work was partially supported by DRET, Publication by courtesy of DRET.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have