Thesis: Decoding Cognitive States under Varying Difficulty LevelsPublished 2 IEEE conference papers, and 3 IEEE journal articles are under review/preparation.Supervised 2 Master-level projects and 1 summer intern.Acted as a TA for "Brain-Computer Interaction'' and "Data Analysis and Model Classification'' classes, each for two semesters.Daily activities:- Design and build protocols for collecting data with Unity (C#) and Python. Including piloting a simulated drone with various sizes of waypoints and capturing auditory effort under different noise.- Record EEG and EOG signals from human subjects with a Biosemi system, and also utilized g.Tec and ANT Neuro for side projects.- Process EEG and EOG signals using multiple signal processing techniques and supervised machine learning methods mainly with MATLAB and call Python functions when needed. To name a few, RPCA, ICA (for removing ocular artifact), regularized GLM regression, LDA.- Conduct closed-loop experiments with real-time decoders.Research detail:Investigated the effect of using EEG signals to automatically optimize the level of difficulty when piloting a simulated drone. This requires an objective method to define subjective difficulty level, for which I used logistic regression. Another important aspect is to build an accurate real-time EEG decoder with limited data (where deep learning cannot apply). I carefully chose different regularized methods and utilized past information to boost accuracy, and meanwhile, the decoder also automatically removes ocular artifacts.The developed decoder has the potential to replace the subjective decision process of difficulty level for the subjects with consistent patterns across different days. The decoder generally can work better with a longer decision time, e.g., 12 seconds. This implies potential benefits on where an instant decision is not necessary. For example, a teacher can be aware that students are experiencing an inappropriate difficulty level.

Researched on acoustic beam-forming techniques. Implemented beam-forming algorithms such as delay and sum, Frost, adaptive Frost, and MVDR with a dual-microphone system.

Published 5 IEEE/ACM papers, 1 IEEE sponsored paper, and 1 workshop paper.Conducted and led research projects as below:- Calibration of EEG signals across days: Enabled benefit of using multiple days of EEG data. EEG signals can distinct on different days even doing the same task. Therefore, even the common sense of machine learning is that using bigger data equals a better model, this did not immediately apply to the targeted EEG data. We proposed to use robust PCA to filter out the components of useful. The U.C. San Diego in the U.S.A. was the research partner.- Brain-Computer Interface based Sound Source Separation: Developed a decoder telling the musical instrument being attended when two instruments are being played. The decoder is used to emphasize the volume of interest.- Sound Source Separation: Improved 2 dB in the source-to-distortion ratio by convolutional sparse coding (CSC) in a setting of score-informed monaural music, and in the case of absence of the score, a multi-pitch estimator can be used and was tested. Many methods focus on decomposing the power spectrum of the mixture without taking care of phase information. On the other hand, the proposed method utilized temporal waveform to decompose. I noticed a fast CSC algorithm and made the collaboration.Demo: http://mac.citi.sinica.edu.tw/research/CSC_separation/- Dictionary-based Music Genre Retrieval System: Accelerated 8X while achieved state-of-the-art performance by a screening method. Sparse representation relies on decomposing music signals into a few defined "musical words". This method, however, benefits from a big dictionary at a cost of long computational time. I employed a mathematical-guaranteed method to speed up.Organized a reliable and high-performance computing environment based on NAS.

Supervised the expense of the company.Completed airborne training.