Investigations into Neurochaos Learning Architectures for Effective Data Classification

Abstract

Artificial neural networks(ANN) are developed to mimic the biological neurons. Actually newlinebrain neurons are inherently nonlinear and found to exhibit chaos. Existing ANN newlinealgorithms are only slightly similar to the actual human brain. The learning ability of newlinebiological neural networks is very high compared to the existing developed ANNs. For newlineexample, a human brain can differentiate between a cat and a dog after seeing only a newlinefew sets of sample images of cats and dogs. However, Deep Learning(DL) architectures newlineneed sufficiently large data samples to learn, predict, and classify correctly. Also unlike ANNs, biological neural networks are more robust to noise and interference. In a biological neuron, chaos is a fundamental property that exhibits at the level of a biological neuron and also at different spatiotemporal scales. Development of artificial neural newlinenetworks based on chaos theory named as Neurochaos Learning (NL) [5, 6] influences newlinethe researchers to analyze the application of chaotic maps as neurons for classification newlinetasks.NL gives comparable performance with state of the art Machine Learning (ML) newlinemethods and sometimes exceeds them, especially in the low training sample regime. newlineNL uses 1D chaotic maps namely Generalized Lur¨oth Series (GLS) as neurons and the newlinesuccess of NL owes to the rich properties of these chaotic GLS maps. In the first phase of newlinethis research, we investigate whether 1D Logsitic map which exhibits rich properties of chaos increase the performance of NL. For that, we developed ChaosFEXLogistic structure newlineand analysed the classification performance for various well known publicly available newlinedatasets such as Iris, Bank Note Authentication, Ionosphere, Wine, Breast Cancer newlineWisconsin, Statlog(Heart), Seeds and Haberman s Survival. In ChaosFEXLogistic newlinearchitecture, the NL features generated are fed to cosine similarity classifier for classification.We also proposed ChaosFEXLogistic +SVM, where the NL features are fed to linear SVM classifier. For Ionosphere, Bank Note Authentication, Haberman s..