Implementation of a Deep ReLU Neuron Network with a Memristive Circuit
Martin Klimo, Peter Tarábek, Ondrej Šuch, Juraj Smieško and Ondrej Škvarek
Convolutional neural networks exhibit state-of-the-art performance in many image recognition tasks. Their software implementation on computers with von Neumann architecture limits the size of tasks that are possible to solve in real time. Therefore, hardware implementations are also important, and memristive circuits are a promising new technology for such hardware implementation.We aim to demonstrate the possibility of the direct transfer of ReLU neural networks trained in software to memristive circuit structure. We use a dendrite spine-like topology to implement weighting coefficients and the division of signal values between positive and negative values using dual structure neural network. The results of the simulation of digit recognition on the benchmark MNIST database confirm the possibility of achieving sufficient precision using 4-bit binary weighting coefficients. Our simulation of the memristive implementation of a neural network assumed ideal memristor elements.
Keywords: Convolutional neural networks, rectified linear unit, resistive switching, memristive circuits, memristive synapse implementation, pattern recognition, image recognition, fuzzy logic.