Practical gradient-descent for memristive crossbars

Abstract

This paper discusses implementations of gradientdescent based learning algorithms on memristive crossbar arrays. The Unregulated Step Descent (USD) is described as a practical algorithm for feed-forward on-line training of large crossbar arrays. It allows fast feed-forward fully parallel on-line hardware based learning, without requiring accurate models of the memristor behaviour and precise control of the programming pulses. The effect of device parameters, training parameters, and device variability on the learning performance of crossbar arrays trained using the USD algorithm has been studied via simulations. Abstract This paper discusses implementations of gradientdescent based learning algorithms on memristive crossbar arrays. The Unregulated Step Descent (USD) is described as a practical algorithm for feed-forward on-line training of large crossbar arrays. It allows fast feed-forward fully parallel on-line hardware based learning, without requiring accurate models of the memristor behaviour and precise control of the programming pulses. The effect of device parameters, training parameters, and device variability on the learning performance of crossbar arrays trained using the USD algorithm has been studied via simulations.There is a significant interest in using memristive devices for computation, in particular in the context of neuromorpic systems [1] and artificial neural networks [2][3][4][5][6][7]. Memristors are typically fabricated in the form of highly-dense crossbar arrays, which naturally lend themselves to the vector-matrix multiplications that are at the core of the neural network algorithms. Memristor-based hardware implementations, while promising low-power high-speed computation, need to address several challenges, such as extreme device variability, complex state-dependent behaviours, or difficulty in integrating active devices within the crossbar. In this paper, we discuss an approximate gradient-descent based learning algorithm, called the Unregulated Step Descent (USD) [8] that addresses these hardware issues, and provides a practical method for training large crossbar arrays in machine learning applications.