Fig. 4. Fashion product and digit recognition using CNNs.

a An experimental setup for implementing the dot-product operation. b A CNN architecture for both fashion product and handwritten digit recognition. Insets of fashion products are shown for illustration purpose. The photonic–electronic dot-product engine performs the convolution for MNIST fashion products⁴⁵ and handwritten digits⁴⁶ with an image size of 14 × 14. c, d Extracted features showing measured results with labelled SDs compared with those of theoretically calculated ones using a 32-bit computer for a T-shirt (c) and digit zero (d). e, f Confusion matrices showing comparable inferencing accuracy for fashion product recognition with 86% from experiment and 87% from calculation (e), and digit recognition with 87% from experiment and 88% from calculation (f).