Neural network image classifiers informed by factor analyzers (Q7007366)

In many real-world machine learning problems, e.g. in computer tomography or satellite images, the number of images is severely limited (less than 1000) by the complexity and cost of obtaining them or the labor intensity of labeling. Here, a method to increase the neural networks (NN) learning rate on such datasets is proposed and tested using factor analyzers (FA) with additive and impulse noise components as additional features of the physical data model; the novelty of the authors' proposal is that the FA with additive and impulse noise is embedded in the NN and trained together with it. The authors prove that the system of random variables of the factor analyzer with additive impulse noise is identifiable and test the method using the Oxford flowers dataset (102 classes, large between class similarity and small within class similarity) and the UC Merced dataset of RGB satellite images of land use (21 classes) using popular NN architectures (EfficientNet B0, MobileNetV2, Xception) with pre-trained weights. The authors claim that for all datasets and NN architectures, their NN is better: the median increase in average classification accuracy across all architectures for the Oxford Flowers set was 1.39\% and for UC Merced it was up by 2.51\%, but the claim is not provable. Programs are not provided. Results in this paper are licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution, and reproduction in any medium or format.