The entire architecture uses a caricature of neurons as binary variables which
are either on or off. This is clearly a very simplistic assumption, but it has
several advantages. The description of the algorithm is greatly simplified, its
serial implementation becomes very efficient, and it does not a-priori commit to
any specific assumption on the information conveyed in the continuous output of
any of the neurons. On the other hand, starting from the model presented here, it
is possible to incorporate multivalued outputs for the neurons, at each stage, and
carefully study how this affects the performance, the stability and the invariance
properties of the system.