NeuroSOFM-Classifier: A Low Power Classifier Using Continuous Real-Time Unsupervised Clustering

Abstract

Supervised machine learning techniques are becoming subject of significant interest in data analysis. However, the high memory bandwidth requirement of current implementations and scarcity of labeled data in many applications prevents implementation of supervised machine learning techniques. In this work, we propose a neuromorphic architecture implementing the self-organizing feature map algorithm using ferroelectric field-effect transistors (Fe-FETs) and gated-resistive random-access memory (gated-RRAM) to produce a semi-supervised NeuroSOFM-Classifier. A best matching input (BMI) identifier circuit allows for very few labeled samples to be used to provide supervised class labels for each neuron in the NeuroSOFM-Classifier. The best matching unit (BMU) or neuron for consequent samples can then be used to inference or classify the new data. This NeuroSOFM-Classifier, trained on just 1% of the labeled data, is capable of classifying COVID-19 patient chest x-rays with 96% accuracy.