Abstract
It has long been recognised that statistical dependencies in neuronal activity need to be taken into account whendecoding stimuli encoded in a neural population. It is far less well appreciated that the same decoding challengesarise in the context of autoassociative memory, when retrieving information stored in correlated synapses. Suchcorrelations have been well documented experimentally (Song et al, 2005); here we show how they can arisebetween synapses that share pre- or post-synaptic partners when any of several well-known additive (Hopfield,1982) or metaplastic (Fusi et al, 2005) learning rules is applied. To assess the importance of these dependen-cies for recall, we adopt the strategy of comparing the performance of decoders which either do, or do not, takethem into account, but are otherwise optimal, showing that ignoring synaptic correlations has catastrophic conse-quences for retrieval. We therefore study how recurrent circuit dynamics can implement decoding that is sensitiveto correlations. Optimal retrieval dynamics in the face of correlations require substantial circuit complexities. Bycontrast, we show that it is possible to construct approximately optimal retrieval dynamics that are biologicallyplausible. The difference between these dynamics and those that ignore correlations is a set of non-linear circuitmotifs that have been suggested on experimental grounds, including forms of feedback inhibition and experimen-tally observed dendritic nonlinearities (Branco et al, 2011). We therefore show how assuaging an old enemy leadsto a novel functional account of key biophysical features of the neural substrate.
