It seems to me that the location inference algorithm on htmresearch has a real limitation in that the network learns to infer objects presented only from a set of discrete body locations (theoretically extracted from hypocampal grid cells activities, I suppose). Do we have an idea of how the size of this set depends on the parameters of the network (i.e. how large can it practically get)? Is biologically the process of generalization across continuous space of body positions (after learning on a sample) job of this network, or of completely another mechanism?
EDIT: Sorry, noticed that the question doesn’t make sense, the network does generalize to any body location, and I can’t delete this question now.