Why do neurons in cortex respond so selectively to words, objects, and faces?

Bowers, J. 1 , Damian, M. 1 , Davis, C. . 2 & Vankov, I. 1

1 University of Bristol
2 University of London, Royal Holloway

A key insight from 50 years of neurophysiology is that single neurons in cortex often respond to information in a highly selective manner. But why? We report a series of simulations that show that PDP networks learn highly selective representations when trained to encode multiple things at the same time in short-term memory. We argue that it is necessary to learn selective codes I this context in order to overcome the superposition catastrophe. That is, co-activating multiple distributed codes leads to a blend pattern that is ambiguous, making distributed codes ill-suited for supporting STM. Selective (even local) coding, by contrast, does not result in ambiguous blends, and accordingly, provides a good medium for STM. Given that multiple cognitive systems support the co-activation of 4+/1 things, this suggests that there is a computational pressure to learn local codes in the cortex.