Neural networks optimization via Gauss–Newton based QR factorization on SARS-CoV-2 variant classification

Abstract

Studies on the COVID-19 pandemic continue due to the potential mutation creating new variants. One response to be aware of the situation is by classifying SARS-CoV-2 variants. Neural networks (NNs)-based classifiers showed good accuracies but are known very costly in the learning process. Second-order optimization approaches are alternatives for NNs to work faster instead of the first-order ones. Still, it needs a huge memory usage. Therefore, we propose a new second-order optimization method for NNs, called QR-GN, to efficiently classify SARS-CoV-2 variants. The proposed method is derived from NNs and Gauss–Newton with QR factorization. The goal of this study is to classify SARS-CoV-2 variants given their spike protein sequences efficiently with high accuracy. In this study, the proposed method was demonstrated on a public dataset for the protein SARS-CoV-2. In the demonstrations, the proposed method outperformed other optimization methods in terms of memory usage and run time. Moreover, the proposed method can significantly elevate the accuracy classification for various NNs, such as: single layer perceptron, multilayer perceptron, and convolutional neural networks.