- This event has passed.
Thesis Defence of Mr. Jerrin Thomas Panachakel
October 11, 2022 @ 5:30 PM - 6:30 PM IST
Degree Registered: Ph.D.
Date and Time: Oct. 11, 2022 (Tuesday) 12 Noon
Venue: MMCR, Hall No. C 241, II Floor, Dept. of Electrical Engineering.
MS Teams Link: https://teams.microsoft.com/l/channel/19%3a7nfRrQdIH7fPf5u003TVCBAp18cRM7dyFwv1eqJRGjY1%40thread.tacv2/General?groupId=53beb192-2a2f-4f20-8a0c-dbb67a65f532&tenantId=6f15cd97-f6a7-41e3-b2c5-ad4193976476
Title: Machine Learning for Decoding Imagined words and Altered State of Consciousness from EEG
Abstract: The thesis explores several architectures for accurately decoding the cognitive activity from EEG recorded during speech imagery and Rajayoga meditation. The major contributions of the thesis are listed below:
Neural Correlates of Phonological Category in Speech Imagery EEG
- We have shown that neural correlates of phonological categories exist in the EEG recorded during speech imagery. These correlates lead to significant differences in the mean phase coherence (MPC) values of the EEG across several cortical regions.
- We have also shown that MPC values can be used for accurately classifying the EEG recorded during speech imagery based on the phonological category of the prompts. The proposed architecture for this task has an accuracy of 84.9%.
Decoding Imagined Words from EEG
- One of the challenges in designing systems for decoding imagined words from EEG is the limited availability of data. We have presented three architectures for decoding imagined words from EEG. All three architectures alleviate this problem of limited availability of data.
- The transfer learning-based architecture employs MPC and magnitude-squared coherence values along with a ResNet50-based classifier. This architecture achieves an accuracy of 92.8% on a publicly available EEG dataset in classifying speech imagery.
Classification of Altered State of Consciousness from Resting State
- We have presented three architectures for classifying the altered state of consciousness during Rajayoga meditation from the resting state.
- Both CSP-LDA-LSTM (common spatial pattern-linear discriminant analysis-long short-term memory) and SVD-DNN (singular value decomposition-deep neural network) architectures are able to capture subject-invariant features.
- The best intra-subject accuracy obtained is 98.2% and the best inter-subject accuracy is 96.4%.