Simplified neurochaos learning architectures for data classification.

Developing machine learning algorithms that can classify datasets with higher accuracy and efficiency is crucial in practical applications. Neurochaos learning (NL) is a recently proposed algorithm that is inspired by the chaotic firing of neurons in the brain. NL has shown promise in recent times both in terms of classification accuracy and in the number of samples needed for training. In this study, we propose a novel simplification of the neurochaos learning algorithm by reducing the number of features needed for classification and also reducing the number of hyperparameters needed to be tuned. By using a single feature of the chaotic neural traces (orbit generated by chaotic map) of NL and by using only one hyperparameter, we demonstrate a significant boost in run time of the algorithm while retaining comparable classification accuracy. This single feature could either be the mean of the chaotic neural traces (Tracemean) or the Fluctuation Index (FI) of the chaotic neural traces. The classifier itself could either be a simple cosine similarity (Tracemean ChaosNet, FI ChaosNet) or any of the classical machine learning (ML) classifiers (Tracemean+ML, FI+ML). We compare the performance of these newly proposed simplified NL algorithms on ten publicly available datasets. The proposed simplified NL architectures in this study are able to efficiently classify datasets while taking much less run time. The fact that only a single hyperparameter needs to be tuned in both architectures (Tracemean ChaosNet and FI ChaosNet) makes them very attractive for practical applications with the ease of interpretability.