NeuroSOFM: A Neuromorphic Self-Organizing Feature Map Heterogeneously Integrating RRAM and FeFET

Abstract

Many currently available hardware implementations of the unsupervised self-organizing feature map (SOFM) algorithm utilize CMOS-only circuits that often compromise key behaviors of the SOFM algorithm due to complexity. We propose a neuromorphic architecture harnessing the unique properties of FeFETs and gated-RRAM for in-memory computing to implement the SOFM algorithm. The FeFET-based synapse, organized in a novel circuit, is able to compute the input-weight Euclidean error in memory via the saturation drain current. The selfdecaying states of the gated-RRAM allow for a self-decaying neighborhood and learning rate implementation to allow for convergence and lifelong learning. This novel architecture is able to successfully cluster benchmarks (RGB colors and MNIST handwritten digits) and real-life datasets such as COVID-19 patient chest x-rays completely unsupervised. The architecture also demonstrates a significant amount of robustness to device variability and damaged neurons. Additionally, the proposed architecture is completely parallelized and provides a power efficient platform for implementing the SOFM algorithm.

Self-Organizing Mapping Neural Network Implementation Based on 3-D NAND Flash for Competitive Learning

“…Implementing SOM with conventional CMOS-based hardware is restricted by the complexity in calculating the ED and winner determination, which imposes an enormous increase in time cost and energy consumption when the number of neurons increases. Until now, several emerging nonvolatile memories, containing resistive random access memory (RRAM) and ferroelectric field-effect transistor (FeFET), were proposed as synaptic devices to implement SOM neuromorphic systems [3]- [7]. However, compared with these devices, 3D NAND Flash has obvious advantages in terms of cost, retention, density, and technology maturity [8]- [10].…”