Fig. 5: Analog near-sensor computing for handwriting recognition.

a The hardware used in this experiment: a handwriting sensor (resistive touch screen), a front-end circuit, and two 4 × 8 eRNR circuits for the x- and y-axes of the sensor. b A conceptual schematic of analog near-sensor computing without any digital memory. The front-end circuit drives the resistive touch screen and allows it to collect the handwriting information, which is then converted into two-dimensional x- and y-analog signals. These signals are then input into two 4 × 8 parallel eRNR circuits. The trained analog weights W_out in the memristor array are used to obtain the classification output for the five handwritten vowels. c–f The signal flows measured from the eRNR hardware for different handwritten patterns, including c the five handwritten vowels, d the sensory signals for the x- and y-axes x(k), e the 64 channel reservoir states s(k) of the eRNRs, and (f) the output y(k) computed based on s(k) and the trained weights. g Confusion matrix using W_out without noise-aware training. The overall accuracy is 97.1%. h Classification accuracy as a function of simulated memristor conductance variation with and without the noise-aware training method. The measured average variation of the memristor array was 0.368 μS. i Confusion matrix using analog W_out stored in the memristor array. The overall accuracy was 94.0%, with a standard deviation of 0.8%.