Towards Engram Memory Machines. A Feasibility Study of Image Classification Using Warping and Low-dimensional Class Prototypes

Abstract

Against the backdrop of ongoing green AI research to reduce the environmental footprint of artificial intelligence technology, a new supervised machine learning approach is presented with potential to contribute to the development of artificial systems capable of learning incrementally from new data. This article reports results from a feasibility study on image classification using the Modified National Institute of Standards and Technology data set. Training consists of producing class-level image prototypes, referred to as engrams, using a single-pass method. At inference stage, previously unseen images are warped to match the engrams individually, using a continuous transformation including a tuneable rigidity parameter. The images are classified by maximising similarity to the engrams across classes following warping. Unlike alternative approaches recording binary feature coincidences in highdimensional spaces, this method operates on low-dimensional data representations. In addition to enhancing interpretability, this has the advantage of circumventing memory scaling bottlenecks. A test accuracy above 90% was observed in this study, which is a promising result considering that the engrams exclusively encoded average data patterns. Future work will focus on reducing inference time, enhancing classification accuracy and estimating the performance of the method on additional datasets.