A Model of Clipped Hebbian Learning in a Neocortical Pyramidal Cell

A detailed compartmental model of a cortical pyramidal cell is used to determine the effect of the spatial distribution of synapses across the dendritic tree on the pattern recognition capability of the neu-ron. By setting synaptic strengths according to the clipped Hebbian learning rule used in the associative net neural network model, the cell is able to recognise input patterns, but with a one to two order of magnitude decrease in performance compared to network computing units. Cell performance is optimised by particular forms of input signal, but is not altered by different pattern recognition criteria. Paradigm We compare the pattern recognition performance of a model pyra-midal cell with that of a computing unit from an associative net. ? Local Hebbian learning allows an output unit in the net to recog-nise input patterns with which it was associated. By setting synaptic strengths in the pyramidal cell dendrites to be the same as the weights onto the output unit, will the cell recog-nise the same input patterns as the unit? The associative net 1 is a simple model of heteroassociative memory. Input and output units with binary activity (0 =inactive, 1 =active) are connected by feedforward synapses that are also binary. Pairs of patterns are stored using a clipped Hebbian learning rule: – the weight of a synaptic connection is changed from 0 to 1 if both input and output units are active for the same pattern pair Output units (B) Input units (A) Matrix of synaptic weights after storage of a large number of pattern pairs – purple: weight 0 – red: weight 1 Recall of a previously stored output pattern: – a stored input pattern is presented on the input units as a cue – each output unit calculates the dendritic sum of its inputs (number of active inputs connected by synapses with weight 1) – a threshold is applied to these dendritic sums to determine which output units should be active – the resultant active outputs constitute the recalled pattern. For a single output unit, memory performance can be gauged by its ability to distinguish between the input patterns with which it was coactive during storage and those with which it was not – this can be measured by the signal-to-noise ratio (see Performance Criterion). We examine output unit performance in the following situation: – the output unit receives connections from 4096 input units – input patterns contain …