This kinda early paper (or it reads like an early paper) shows how easy it is to train multilayer winner-takes-all neural networks. It greatly simplifies things because you only need to reward or punish a limited number elements in the net that have won. And if an element is punished enough an alternative will spring up which will change the behavior of the network (causing exploration of a different behavior).
http://mocha-java.uccs.edu/dossier/RESEARCH/2001annie-.pdf
From reading some of the papers from Numenta it seems that winner-takes-all mechanisms are important in the human brain.
1 Like