Fig. 4: Implementation of Heterosynaptic Modulation for Enhanced MNIST Digit Recognition.

a Schematic representation showing how consideration of MNIST digit datasets, in this case downscaled (8 × 8-pixel) and binarized “5’s” and “8’s”, can lead to identification of areas of greater and lower importance to digit recognition, showcased here as areas of greater and lower cumulative pixel intensity (Δ pixel intensity), respectively, when images are overlapped. b, c By taking the difference in Δ pixel intensity between two digits of interest, (b) “5” and “8” and (c) “8” and “9”, the features of highest and lowest contrast may be identified. d, e Confusion matrices between (d) “5” and “8” and (e) “8” and “9” for the hardware-based inference performed on a 64 × 10 sub-section of the array in Fig. 3. f, g Confusion matrices between (f) “5” and “8” and (g) “8” and “9” for cases wherein heterosynaptic modulation is dynamically applied to areas of high and low contrast between digits, as identified in (b) and (c), respectively. Increases in overall classification accuracy of 14.5% and 4% can be seen for “5” and “8” and for “8” and “9”, respectively, compared to the standard case.