Hardware Implementation of nmODE on FPGA

Neural Ordinary Differential Equations (ODEs) have attracted much attention since their introduction. They use a black-box differential equation solver instead of iterating repeatedly to calculate the network output, which makes a tradeoff between numerical precision and speed. The research on ODE mainly focused on theoretical orientation, and lacks practical applications, especially in hardware deployment. In this paper, we propose a hardware implementation of the neural memory ODE model based on Field-Programmable Gate Arrays (FPGAs). Specifically, we propose a neural memory ODE (nmODE) numerical mapping method to accelerate its differential equation solver speed and save memory consumption. After that, we design a nmODE neuron core on FPGA and expand it to support parallel computation. This work is implemented on Xilinx XC7z030 and evaluated with the MNIST dataset. The experimental results demonstrate that nmODE on FPGA achieves a comparable accuracy of 98.84% compared to the baseline nmODE on GPU while accelerating the inference speed by x31 times.