Parallel synapses with transmission nonlinearities enhance neuronal classification capacity
Fig 2
Increased memory capacity with parallel synapses.
(a & d). Illustration of binary classification task. In a two-dimensional input space, four random patterns are labeled with + 1 (white dots) or –1 (black dots) labels. The goal is to find a decision boundary (dashed line) that correctly classifies all data points. In (a), the lower left data point with + 1 label is misclassified, and thus the classification problem has not been solved yet. In (d), all data points are correctly classified, and thus the problem has been solved successfully. (b). The capacity P*/N of a neuron with different numbers of input axons and different numbers of parallel synapses. With increasing N, the capacity keeps increasing, and thus P* grows faster than linearly. (c). Examples of learned effective synaptic transmission functions, normalized by their maximum value. Here, N is 1000, M is 10 and P is 11000. (e). Histogram of amplitude values for all synapses in (c). (f). Histogram of threshold values for all synapses in (c). (g). The relationship between parallel synapses’ amplitudes and their threshold values. The thresholds for all parallel synapses in (c) are divided into the same bins as in (f). We plot the average amplitude of the parallel synapses with thresholds in each bin.