The Neural Heat Exchanger

Jürgen Schmidhuber
IDSIA

In S. Amari, L. Xu, L. Chan, I. King, K. Leung, eds., Progress in Neural Information Processing: Proceedings of the Intl. Conference on Neural Information Processing, pages 194-197, Springer, Hongkong, 1996. (Invited Talk; earlier presentations in talks at various universities since 1990.)

Abstract:

The ``Neural Heat Exchanger'' is an alternative, supervised learning method for multi-layer neural nets. It is inspired by the physical heat exchanger. Unlike backprop, it is entirely local. This makes its parallel implementation trivial. It was first presented during occasional talks since 1990, and is closely related to Hinton et. al.'s recent Helmholtz Machine (1995). For the first time, this paper presents the basic ideas in written form. To fully understand the Neural Heat Exchanger's advantages and limitations, however, much theoretical and empirical work remains to be done.