Multimodal transistors as ReLU activation functions in physical neural network classifiers

Isin Surekcigil Pesch,Olivier De Sagazan,Radu A Sporea,Adnan Mehonic,Eva Bestelink

doi:10.1038/s41598-021-04614-9

Abstract

Artificial neural networks (ANNs) providing sophisticated, power-efficient classification are finding their way into thin-film electronics. Thin-film technologies require robust, layout-efficient devices with facile manufacturability. Here, we show how the multimodal transistor’s (MMT’s) transfer characteristic, with linear dependence in saturation, replicates the rectified linear unit (ReLU) activation function of convolutional ANNs (CNNs). Using MATLAB, we evaluate CNN performance using systematically distorted ReLU functions, then substitute measured and simulated MMT transfer characteristics as proxies for ReLU. High classification accuracy is maintained, despite large variations in geometrical and electrical parameters, as CNNs use the same activation functions for training and classification.

Highlights

Current systems using CMOS, digital technologies with von Neumann architectures, are not best suited to support a massive increase in computing power demands driven by AI development[1,2,3]
The main challenge restricting costeffective development is the thin-film transistor (TFT), a device that comprises the backbone of many large area electronics (LAE)
We investigate the practicality of using the multimodal transistor (MMT’s) transfer characteristic as a viable rectified linear unit (ReLU) activation functions (AFs) for future thin-film artificial neural networks (ANNs) with high classification accuracy, despite relatively large process variations expected in such technologies

Summary

Introduction

Current systems using CMOS, digital technologies with von Neumann architectures, are not best suited to support a massive increase in computing power demands driven by AI development[1,2,3]. While uniformity of operation is a requirement for array-based L AE12, this has not limited the interest of exploring TFTs in edge processing alongside other thin-film architectures[3,14,15,16] In this context, analog implementations of signal processing functions are attractive, especially if the TFTs utilized are energy-efficient, as well as robust against variations during manufacturing and operation[12]. The multimodal transistor (MMT)[19] (Fig. 1a, b) is a TFT with superior functionality, robustness and energy efficiency, especially in analog and mixed-signal applications It can be designed with a linear dependence between input voltage and output current even when operating in s aturation[19], making it highly suited to operation as a rectified linear unit (ReLU)[20], as this function is immediately achievable[19]. Would be highly beneficial for further development of ANN accelerators based on non-CMOS analog devices and in-memory computing concepts

Methods

Results

Discussion

Conclusion