THE REPRESENTATIONAL CAPACITY OF THE DISTRIBUTED ENCODING OF INFORMATION PROVIDED BY POPULATIONS OF NEURONS IN PRIMATE TEMPORAL VISUAL-CORTEX

It has been shown that it is possible to read, from the firing rates o f just a small population of neurons, the code that is used in the mac aque temporal lobe visual cortex to distinguish between different face s being looked at. To analyse the information provided by populations of single neurons in the primate temporal cortical visual areas, the r esponses of a population of 14 neurons to 20 visual stimuli were analy sed in a macaque performing a visual fixation task. The population of neurons analysed responded primarily to faces, and the stimuli utilise d were all human and monkey faces. Each neuron had its own response pr ofile to the different members of the stimulus set. The mean response of each neuron to each stimulus in the set was calculated from a fract ion of the ten trials of data available for every stimulus. From the r emaining data, it was possible to calculate, for any population respon se vector, the relative likelihoods that it had been elicited by each of the stimuli in the set. By comparison with the stimuli actually sho wn, the mean percentage correct identification was computed and also t he mean information about the stimuli, in bits, that the population of neurons carried on a single trial. When the decoding algorithm used f or this calculation approximated an optimal, Bayesian estimate of the relative likelihoods, the percentage correct increased from 14% correc t (chance was 5% correct) with one neuron to 67% with 14 neurons. The information conveyed by the population of neurons increased approximat ely linearly from 0.33 bits with one neuron to 2.77 bits with 14 neuro ns. This leads to the important conclusion that the number of stimuli that can be encoded by a population of neurons in this part of the vis ual system increases approximately exponentially as the number of cell s in the sample increases (in that the log of the number of stimuli in creases almost linearly). This is in contrast to a local encoding sche me (of ''grandmother'' cells), in which the number of stimuli encoded increases linearly with the number of cells in the sample. Thus one of the potentially important properties of distributed representations, an exponential increase in the number of stimuli that can be represent ed, has been demonstrated in the brain with this population of neurons , When the algorithm used for estimating stimulus likelihood was as si mple as could be easily implemented by neurons receiving the populatio n's output (based on just the dot product between the population respo nse vector and each mean response vector), it was still found that the 14-neuron population produced 66% correct guesses and conveyed 2.30 b its of information, or 83% of the information that could be extracted with the nearly optimal procedure. It was also shown that, although th ere was some redundancy in the representation (with each neuron contri buting to the information carried by the whole population 60% of the i nformation it carried alone, rather than 100%), this is due to the fac t that the number of stimuli in the set was limited (it was 20), The d ata are consistent with minimal redundancy for sufficiently large and diverse sets of stimuli. The implication for brain connectivity of the distributed encoding scheme, which was demonstrated here in the case of faces, is that a neuron can receive a great deal of information abo ut what is encoded by a large population of neurons if it is able to r eceive, its inputs from a random subset of these neurons, even of limi ted numbers (e.g. hundreds).