Neural networks based on memristor crossbars and bridges

5^th International Conference and Exhibition on Lasers, Optics & Photonics

November 28-30, 2016 Atlanta, USA

Mikhail Tarkov

A V Rzhanov Institute of Semiconductor Physics - RAS, Russia

Posters & Accepted Abstracts: J Laser Opt Photonics

Abstract :

Hardware implementation of a neural network requires a lot of memory for storing the weight matrix of the neurons layer and is expensive. Solving this problem is simplified by using device as the memory cell called memristor. The memristor has many advantages such as non-volatile storage media, low power consumption, high density integration and excellent scalability. Unique ability to retain traces of the device excitation makes it an ideal candidate for the implementation of electronic synapses in neural networks. The present contribution reports for programming memristor array (crossbar). An algorithm for mapping weight matrix of the neuron layers onto memristor crossbar is proposed. LTSPICE model of WTA (â�?�?Winner Takes Allâ�?) neural network implementation on the memristor crossbar and CMOS transistors is developed for binary images recognition. The resistor bridges containing memristors are used for LTSPICE simulation of tunable weights in electronic associative memories, a bi-directional associative memory and an associative memory based on the Hopfield network, which can be implemented as networks of coupled phase oscillators. The experiments using LTSPICE models show that for the reference binary images with size 3x3 the oscillatory Hopfield network converges to the reference images (accordingly, to their inversion) with a random uniform distribution of the binary pixel values in the input images.

Biography :

Email: tarkov@isp.nsc.ru