Fig. 4: bP-PPT array as an optoelectronic CNN for image recognition.
From: Programmable black phosphorus image sensor for broadband optoelectronic edge computing
a A CNN model for classifying handwriting numbers “0” and “1” from the MNIST dataset. The CNN consists of two convolution kernels, an average pooling layer, and a fully connected (FC) layer. The images captured by the bP-PPT array are further processed by the bP-PPT array in the electrical domain. b The 3 × 3 bP-PPT array is programmed with 5-bit precision to represent two kernels generated by offline training. c The experimental results (top, red) for image recognition using the bP-PPT array are compared with the simulation results (bottom, blue). Each bar is the score indicating the possibility of the CNN recognizing an image in the MNIST image library. The incorrectly recognized cases are in gray color. d The experimental and simulated confusion table for 100 images from the MNIST dataset. Colored diagonal elements in the table indicate the correctly identified cases.
