Fig. 5: 10-class digit recognition.

a Images of the 10 digits in the training set. b Example of the conversion from a 5 × 3 image to 5 3-timeframe pulse trains. c Read current evolutions of a typical volatile FD under 7 different pulse trains representing the pixel arrangements of “111”, “110”, “101”, “100”, “011”, “010”, and “001”. d Experimentally measured reservoir states after stimulating the 5-volatile-FD reservoir with different input images in the training set. e Images of the 12 digits in the test set, which are the noisy versions of their corresponding training samples in (a). The noisy “9” is incorrectly recognized as “8” (marked in red), whereas all the rest digits are correctly recognized.