Session: SYMP 6-10: Reservoir Computing and Control

Paper Number: 140361

140361 - Brain Inspired Heterogeneous Reservoir Computing With Ion Channel-Based Memristors 

As the demand for energy-efficient computational devices increases, physical reservoir computing (PRC) emerges as a promising solution, especially for edge time-series classification and prediction tasks. PRC is based on software reservoir computing (RC) principles, where a randomly generated "reservoir" layer produces a high-dimensional projection of an input time series. This projection is then used to train the output layer's weights through linear regression, making RC highly efficient compared to conventional recurrent neural networks during the training phase. Both RC and PRC rely on the principle of "fading memory," where the dynamics of the reservoir layer depend more on recent inputs than inputs from the far past. For this reason, dynamic memristors are prime candidates for achieving this property in analog implementations. These devices update their conductance based on present and past voltage stimulation, but unlike their non-volatile counterparts, they return to some base-level conductance after some time with an absence of stimulation. By encoding input data as voltage pulses and measuring the current output, the conductance state of the memristors can be calculated. Previous implementations involved many nominally identical devices running in parallel while employing varied encoding techniques or device-to-device variations to create the high-dimensional projection space. However, to avoid issues arising from complex pre-processing of data, synchronization between encodings, and reliance on stochastic differences in devices, memristors with intentionally varied dynamics are preferred. Each intentionally varied memristor responds uniquely to a single data encoding, producing linearly independent outputs without having to generate individualized encodings for each device. Ion-channel-based memristors with voltage-dependent dynamics can be controllably and predictively tuned through voltage offset or adjustment of the ion channel concentration to exhibit diverse, dynamic properties. Operating natively on biological timescales, these devices are well-suited to bio-information processing tasks. Through experiments and simulations, we have shown that reservoir layers constructed with a small number of distinct memristors exhibit significantly higher predictive and classification accuracies with a single data encoding compared to relying exclusively on stochastic variations in device properties. In a second-order nonlinear dynamical system prediction task, the varied memristor reservoir experimentally achieved a normalized mean square error of 0.0015, using as few as five distinct memristors. Similarly, in a neural activity classification task, a reservoir of just three distinct memristors experimentally attained an accuracy of 96.5%. This novel approach lays the groundwork for next-generation computing systems that can exploit the complex dynamics of their diverse building blocks to achieve increased signal processing capabilities with a minimal amount of pre-processing.

Presenting Author: Nicholas Armendarez Pennsylvania State University

Presenting Author Biography: Nicholas Armendarez is a PhD candidate at Pennsylvania State University. He holds a B.S. in Bio-Systems Engineering from The University of Kentucky. His research focuses on bioinspired smart materials with applications in energy-efficient analog implementations of machine learning algorithms.

Authors:

Nicholas Armendarez
Ahmed Mohamed
Anurag Dhungel
Md Sakib Hasan
Joseph Najem