This is only about spatial pooling.
If I understand correctly, a macrocolumn contains about 100 minicolumns, each being tuned to some preferred stimulus in its receptive field (RF). Neighboring macrocolumns do not overlap in their RF. This in turn means that there are only about 100 feature detectors for a given RF. If those feature detectors implement a winner-takes-all competition between them, that would only allow 100 features to be learnt and recognized for a given RF. That doesn’t sound like a while lot, considering the diversity of inputs for natural stimuli and the precision of recognition (lets say for vision in V1). Possibly they instead use a N-out-of-100 coding which would increase capacity somewhat, but still…
Any thoughts on this anyone ?