Accelerated Neural Network Training through Dimensionality Reduction for High-Throughput Screening of Topological Materials

TL;DR

This paper explores using dimensionality reduction to improve neural network training efficiency for predicting topological material properties, aiming to reduce computational costs while maintaining accuracy.

Contribution

It introduces a method combining dimensionality reduction with AdaBoost to efficiently select features and predict topological properties of materials.

Findings

Dimensionality reduction reduces training time significantly.

Reduced feature sets maintain comparable accuracy.

AdaBoost outperforms neural networks in training speed.

Abstract

Machine Learning facilitates building a large variety of models, starting from elementary linear regression models to very complex neural networks. Neural networks are currently limited by the size of data provided and the huge computational cost of training a model. This is especially problematic when dealing with a large set of features without much prior knowledge of how good or bad each individual feature is. We try tackling the problem using dimensionality reduction algorithms to construct more meaningful features. We also compare the accuracy and training times of raw data and data transformed after dimensionality reduction to deduce a sufficient number of dimensions without sacrificing accuracy. The indicated estimation is done using a lighter decision tree-based algorithm, AdaBoost, as it trains faster than neural networks. We have chosen the data from an online database of topological materials, Materiae. Our final goal is to construct a model to predict the topological properties of new materials from elementary properties.