Neuromorphic Circuit Simulation with Memristors: Design and Evaluation Using MemTorch for MNIST and CIFAR

TL;DR

This paper investigates the use of memristors in neuromorphic circuits for neural networks, demonstrating their feasibility and minimal precision loss through simulations with Memtorch on MNIST and CIFAR datasets.

Contribution

It introduces a methodology for converting digital CNNs into memristive systems using Memtorch and evaluates their performance and non-idealities.

Findings

Minimal inference precision loss (~1%) observed in simulations

Impact of tile size and non-idealities analyzed

Memristors show promise for energy-efficient neuromorphic computing

Abstract

Memristors offer significant advantages as in-memory computing devices due to their non-volatility, low power consumption, and history-dependent conductivity. These attributes are particularly valuable in the realm of neuromorphic circuits for neural networks, which currently face limitations imposed by the Von Neumann architecture and high energy demands. This study evaluates the feasibility of using memristors for in-memory processing by constructing and training three digital convolutional neural networks with the datasets MNIST, CIFAR10 and CIFAR100. Subsequent conversion of these networks into memristive systems was performed using Memtorch. The simulations, conducted under ideal conditions, revealed minimal precision losses of nearly 1% during inference. Additionally, the study analyzed the impact of tile size and memristor-specific non-idealities on performance, highlighting the practical implications of integrating memristors in neuromorphic computing systems. This exploration into memristive neural network applications underscores the potential of Memtorch in advancing neuromorphic architectures.