# A memristive synaptic circuit and optimization algorithm for synaptic control

## TL;DR

This paper proposes a memristive synaptic circuit and an optimization algorithm to enable linear weight control in memristor-based machine learning systems.

## Contribution

A novel optimization method using the artificial bee colony algorithm to determine circuit parameters for linear memristive weight control.

## Key findings

- Weight control was achieved with a mean square error of 2.33×10−4.
- The software-based test accuracy tracking rate reached 98.186%.
- The method enables linear control for online training with any memristor element.

## Abstract

In order for the backpropagation training method, which is widely used for machine learning inference layer, to be directly applied to memristor crossbar arrays, either the weight change must be linear, or since the memristance change is not constant over time, the current memristance value must be kept in memory or changes must be controlled with an algorithm suitable for the used memristance function. To overcome the memory and energy drawbacks of this non-linearity, in this study, the parameters of a memristive circuit that can implement positive and negative weights were determined by the optimization method, using two charge-controlled mathematial memristor equations and a flux-controlled memristor emulator previously defined in the literature. In this way, the simplest linear control of weight change is achieved. Using the artificial bee colony algorithm, the passive element values of a circuit that can perform weight control up to 0.02 sensitivity and the duration of the applied control signal were determined. According to the experimental study, it was seen that weight control was achieved with a mean square error of 2.33\documentclass[12pt]{minimal}
\usepackage{amsmath}
\usepackage{wasysym}
\usepackage{amsfonts}
\usepackage{amssymb}
\usepackage{amsbsy}
\usepackage{mathrsfs}
\usepackage{upgreek}
\setlength{\oddsidemargin}{-69pt}
\begin{document}$$\times $$\end{document}×10−4. Also the tracking rate of software-based test accuracy is 98.186%. With the proposed optimization method and cost function, linear control can be achieved by determining the parameters needed for online training with any memristor element.