Neuromorphic learning machine based on stochastic reservoir computing for time series data processing and classification

Saw, Chia Yee (2023) Neuromorphic learning machine based on stochastic reservoir computing for time series data processing and classification. Masters thesis, Universiti Teknikal Malaysia Melaka.

	Text (24 Pages) Neuromorphic learning machine based on stochastic reservoir computing for time series data processing and classification.pdf - Submitted Version Download (333kB)
	Text (Full Text) Neuromorphic learning machine based on stochastic reservoir computing for time series data processing and classification.pdf - Submitted Version Restricted to Registered users only Download (5MB)

Abstract

Neuromorphic computing is a potential alternative to conventional von Neumann computers for specialised sensory-processing or classification applications. Neuromorphic systems replicate the biophysics of neurobiological networks by replicating the information processing mechanism of biological neurons and synapses, resulting in high connection and parallelism on a smaller footprint. These characteristics contribute to the implementation of neuromorphic architectures in hardware development for time series data classification and recognition. However, hardware implementation of neural network suffers from resource constraints because to the typically large number of nodes utilised in RC networks and the high chip area required for each processing node. Moreover, a high number of features in time series data classification introduces computational burden and leads to excessive hardware calculation overhead. In this work, a reservoir computing based stochastic spiking neural network (SSNN) has been proposed for time series data processing and classification, enabling a more efficient hardware implementation with low computation overhead caused by minimum extracted features. The proposed neuron reservoir is implemented in two applications including ventricular heartbeat classification and human activity recognition (HAR). The 43 recordings of Electrocardiogram (ECG) signals that included both normal and arrhythmic beats from MIT-BIH arrhythmia database obtained from Physio-Net were used in this work for heartbeat classification. Baseline drift and power line interference that are frequently emphasized in ECG readings are minimized by signal denoising. The single feature, QRS complexes, was extracted and fed into the neural reservoir with 20 neurons in cyclic topology for arrhythmias’ similarity calculation and classification. The HAR is evaluated in order to further validate the proposed SSNN approach. This work proposes feature extraction based on subcarrier correlation and pseudocolor variations caused by human movements depicted in the images using convolution neural network (CNN) without preprocessing applied and enabling low computational complexity and visual observation of entire pattern changes. The extracted features are fed into the neural reservoir for activities recognition. A two-input stochastic neuron is developed for complex machine learning using the stochastic computing (SC) theory. The 20 stochastic neurons are then arranged into a simple cycle reservoir (SCR) architecture to create the SSNN. The proposed system has been implemented in Xilinx Zynq-7000 field-programmable gate array (FPGA) to demonstrate the hardware efficiency leads by the minimum feature size used. The proposed stochastic spiking reservoir achieves an accuracy of 96.91% in heartbeat classification and 92.94% and 93.91% for features based on subcarrier correlation and pseudocolor plot in HAR, demonstrating that the system is accurate and effective at classifying time series data.

Item Type:	Thesis (Masters)
Uncontrolled Keywords:	Neuromorphics, Neural networks (Computer science), Neural computers
Subjects:	T Technology > T Technology (General) T Technology > TA Engineering (General). Civil engineering (General)
Divisions:	Library > Tesis > FKEKK
Depositing User:	Unnamed user with email hanisufiyah@gmail.com
Date Deposited:	12 Nov 2024 10:19
Last Modified:	12 Nov 2024 10:19
URI:	http://eprints.utem.edu.my/id/eprint/27733
Statistic Details:	View Download Statistic

Actions (login required)

View Item