Probing Mus silicium

A nervous system rapidly learns to recognize spoken words, performing correctly for previously unheard examples and speakers. Scientists probe the system (composed of several hundred neurons arranged in two layers) while exposing it to various sounds and performing extracellular recordings on individual cells. Others struggle to gain a global understanding of the mechanism behind its remarkable performance, but two scientists already know the system's secret… because they created it! John Hopfield and Carlos Brody constructed the artificial neural network (playfully dubbed Mus silicium or the sand mouse) to embody a new, robust and biologically plausible principle for categorizing temporal patterns. Instead of publishing it in the traditional way, they announced a contest: could their colleagues deduce the principle by using a traditional anatomical description and physiological experiments on the artificial system? A first paper [Hopfield, J.J. and Brody, C. (2000) Proc. Natl. Acad. Sci. U. S. A. 97, 13919–13924] described the anatomy and physiology of their system; in addition, a web site (http://str.princeton.edu/mus/Organism/) was set up to which contestants could upload sound ‘stimuli’ and then download the ‘response’ of any individual neuron. Three months later, after the contest deadline, Hopfield and Brody published the answer [Proc. Natl. Acad. Sci. U. S. A. (2001) 98, 1282–1287]. The principle, which might be applicable to many different types of temporal sequence learning, is based on the synchrony of neurons with varying decay rates that respond to different features.