Design of Max Pooling Operation Circuit for Binarized Neural Networks Using Single-Flux-Quantum Circuit

Abstract

For the binary neural network (BNN), We design a max pooling operation circuit (MPOC) using single-flux-quantum circuits based on the multiple data comparators. To continuously compare multiple data and find the maximum value, the multiple data comparator uses the internal state of non-destructive read-out (NDRO) flip-flop and NDRO with complementary outputs to store the maximum value. The MPOC is designed with a gate-level pipeline architecture and uses the internal state to store the maximum value, avoiding the use of feedback loops or tree structures when comparing multiple data. Therefore, the area of the MPOC is reduced by approximately 74% compared to the conventional MPOC using many shift registers with tree structures. The throughput is approximately 21.2 times higher and the area is reduced by 22.1% when compared with the conventional MPOC using feedback loops. We designed an MPOC with 2878 Josephson junctions using a 10 kA/cm <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$^{2}$</tex-math></inline-formula> Nb process. The proposed circuit can be used for pooling operations of any size after a 1-channel 3x3 convolutional layer, which accelerates the forward propagation process of binary neural networks (BNNs). Simulation results show that the MPOC can perform pooling operations continuously at the frequency of 50 GHz.