Fig. 5.

a Representative examples of spike raster plots comparing spiking activity in the three model layers for 30 data presentations at three stages; (i) early training, (ii) late training and (iii) testing. Training and test datasets were drawn randomly from the corresponding MNIST datasets but presented stratified in class order. A perfect “stepped” pattern in the output layer thus represents correct classification. The fourth raster plot (iv) shows a 250 ms representative detail from late training covering 3 input presentations. The red and blue superimposed arrows highlight the CN cluster responses to the second and third digit presentations included in the plot. b Comparing classifier performance for the previous model (Diamond et al. 2016a) (offline neural gas = blue dashed trace) and new model (online STDP self-organization = red trace) as the specified number of cluster points (aka virtual receptors) is varied. Ten thousand digits (training 1600 $\times$ 5, test 400 $\times$ 5) were drawn randomly from the corresponding MNIST datasets. For each of 5 runs, digits were presented stratified in class order. Performance is plotted as the average percentage correctly classified according to maximum output spiking activity across the 5 runs. Error bars indicate the standard deviation. The CN spike limit imposed on training was set at 20 spikes maximum, and CN cluster size was set at 30 neurons. c Test regime as in b, but investigating the performance impact of the CN spike limit imposed on training digit presentation in the new model. One hundred cluster points were specified, and CN cluster size was set at 30 neurons. d Test regime as in b, but investigating the performance impact of the CN cluster size employed in the new model. One hundred cluster points were specified, and the CN spike limit imposed during training was set at 20 spikes maximum