Efficient implementation of convolutional neural networks on FPGA

Cyber-Physical Systems (CPS) apply complex machine learning algorithms to process incoming sensor signals to interact with the embedding environment through actuators, which is an inherently real-time task. Therefore, machine learning algorithms must be also evaluated real-time, bringing up very delicate requirements about implementations. Convolutional Neural Networks (CNNs) are artificial neural networks successfully applied in machine vision applications, therefore, their use in CPS is primarily determined by the efficiency of the implementation. CNNs are typically implemented using general purpose CPUs, DSPs, and GPGPUs, however, due to their structure and operation, they are also very good candidates to FPGA implementations. The paper will present an efficient implementation of CNNs on FPGAs and an initial evaluation of the implementation using the LeNet network.