Realizing Linear Synaptic Plasticity in Electric Double Layer-Gated Transistors for Improved Predictive Accuracy and Efficiency in Neuromorphic Computing

TL;DR

This paper demonstrates that linear ionic weight updates in electric double layer-gated transistors improve neuromorphic computing accuracy and efficiency, with a predictive model enabling better training performance on MNIST.

Contribution

The study introduces a predictive linear ionic weight update solver (LIWUS) for EDLTs, enabling linear plasticity in neuromorphic devices and improving neural network training.

Findings

Linear weight updates reduce training epochs by 19%.

Achieved 97.6% accuracy on MNIST with linear updates.

Linear updates outperform nonlinear ones in accuracy and stability.

Abstract

Neuromorphic computing offers a low-power, parallel alternative to traditional von Neumann architectures by addressing the sequential data processing bottlenecks. Electric double layer-gated transistors (EDLTs) resemble biological synapses with their ionic response and offer low power operations, making them suitable for neuromorphic applications. A critical consideration for artificial neural networks (ANNs) is achieving linear and symmetric plasticity (weight updates) during training, as this directly affects accuracy and efficiency. This study uses finite element modeling to explore EDLTs as artificial synapses in ANNs and investigates the underlying mechanisms behind the nonlinear plasticity observed experimentally in previous studies. By solving modified Poisson-Nernst-Planck (mPNP) equations, we examined ion dynamics within an EDL capacitor & their effects on plasticity, revealing that the rates of EDL formation and dissipation are concentration-dependent. Fixed-magnitude pulse inputs result in decreased formation & increased dissipation rates, leading to nonlinear weight updates and limits the number of accessible states and operating range of devices. To address this, we developed a predictive linear ionic weight update solver (LIWUS) in Python to predict voltage pulse inputs that achieve linear plasticity. We then evaluated an ANN with linear and nonlinear weight updates on the MNIST classification task. The LIWUS-provided linear weight updates required 19% fewer epochs in training and validation than the network with nonlinear weight updates to reach optimal performance. It achieved a 97.6% recognition accuracy, 1.5-4.2% higher than with nonlinear updates and a low standard deviation of 0.02%. The network model is amenable to future spiking neural network applications and the performance improvements with linear weight update is expected to increase for complex networks.