Design and Implementation of a Scalable Neuromorphic Classifier for Emotion Detection using EEG Data

Abstract

Emotion classification using EEG signal processing has the potential of significantly improving the social integration of patients suffering from neurological disorders such as Amyotrophic Lateral Sclerosis (ALS) or the acute stages of Azlheimers disease. Emotion classifiers have historically used software installed on general-purpose computers and operating under off-line conditions. Yet the wearability of such classifiers requires the use of low-power hardware accelerators that would enable near real-time classification and extended periods of operations. In this thesis, we architect, design, implement, and test in hardware a neuromorphic processor for emotion classification using a pre-trained Convolutional Neural Network (CNN) that uses de-noised and pre-processed EEG signals as inputs. The EEG signals are generated using a low-cost offthe-shelf device, namely, Emotiv Epoc plus. For CNN training and testing, three repositories of emotion classifications datasets have been combined, including DEAP, DREAMER and an original dataset collected at Khalifa University from 5 healthy subjects using the standard IAPS visual stimulus. The subject-independent CNN classifier has been implemented using the FPGA Atlys board with training provide according to Googles Tensorflow. One important byproduct of our work is a repository of EEG frequency data for emotion detection that is the largest ever. It comprises 60 subjects. Our own collection of data from human subjects has been conducted under the approval of the Committee on Human Subjects Research Ethics at Khalifa University of Science and Technology.

Date of Award	Dec 2017
Original language	American English