Applications of superlattice-like PCM synapses in CNN. (a) Schematic diagram of a CNN comprising feature extraction and classification for a handwritten digit recognition task. Both the convolutional kernels in the feature extraction unit and the connections in the classification unit are simulated by superlattice-like PCM synapses. (b) Output results of the convolutional layer that performs feature extraction on the input picture. (c) The final network recognition rate of the software synapse and synapse data in Fig. 2 using two different update methods.