Abstract
How are we able to recognize visual stimuli that we have not seen for decades? This ability poses a clear challenge to current models of how information is stored in neural circuits. I propose that such long-term memories may depend on the existence of highly selective cortical neurons (effectively grandmother cells) that are so selective that they will not fire at all unless something that closely resembles the original stimulus is seen again. Simulation studies using a simple Spike-Time Dependent Plasticity rule (STDP) have demonstrated that repeatedly presenting a spatiotemporal pattern of spikes will concentrate high synaptic weights on the inputs that fire first during the pattern (Masquelier, Guyonneau, & Thorpe (2008), PLoS ONE, 3, e1377). This process may only need a few tens of presentations to take effect, a suggestion supported by recent work on auditory noise learning (Agus, Thorpe, & Pressnitzer (2010), Neuron, 66, 610). Once the high synaptic weights have been concentrated in this way, the addition of an inhibitory circuit that keeps the total number of active input units under strict control results in a situation where the probability of the cell firing with random input patterns can be arbitrarily low. In effect, the neuron has become extremely selective. Furthermore, if STDP is true, a cell that never fired would be able to maintain their pattern of connections intact for very long periods of time. The existence of cells with such extremely low firing rates is clearly a matter of conjecture, but the possibility that a substantial percentage of cortical neurons could constitute a form of Neocortical Dark Matter, effectively invisible to the neurophysiologist electrode, is one that merits to be taken seriously.