Closed-loop neurofeedback training of attention with real-time EEG
Master's thesis conducted at the Section for Cognitive Systems (advisor: Professor Lars Kai Hansen and Professor Troels Wesenberg Kjaer), Department of Applied Mathematics and Computer Science, Technical University of Denmark (DTU).- Developed a Python-based framework for closed-loop, real-time neurofeedback in EEG. The toolbox contains integration of ultra-precise EEG data sampling with stimuli presentation, subject-specific artifact rejection, cognitive state classification and continuous feedback.
- Implemented a robust real-time classification algorithm for data processing in temporal bins using machine learning techniques.
- Investigated the effect of neurofeedback to modulate top-down attentional states during an attention training task.
- Visualized where discriminative information of decoded attentional states resides in the EEG signature using sensitivity mapping. The aim was to shed light on the temporal dynamics of subjective attentional states.
Poster presented at FENS Brain Conference Dynamics of the brain: Temporal aspects of computation in Rungsted, Denmark, 9-12 June 2019: Tuckute, G., Hansen, S.T., Kjaer, T.W., Hansen, L.K.: Can we control attention with closed-loop EEG neurofeedback? Poster can be found here.
GitHub toolbox (also containing sample data and video demo): Closed-Loop Neurofeedback in EEG using Python.
Multiple projects within neural computations underlying language processing
Work conducted as a full-time research assistant in Ev Fedorenko's Language Lab, McGovern Institute for Brain Research, Massachusetts Institute of Technology (MIT) in June 2018 - September 2018 (and ongoing):- Performed processing and first level analyses of fMRI data using Freesurfer, MATLAB, Linux shell scripting and high-performance computing. Wrote custom MATLAB scripts for second level analyses. Acquired fMRI data.
- Investigated whether anatomical brain properties respect functional network boundaries by analyzing cortical thickness values and gray matter intensities across two separate brain networks: the fronto-temporal language system and the fronto-parietal multiple demand (MD) system.
- Performed seed-based functional connectivity analysis on resting-state fMRI data for implementation of a convolutional neural network (CNN) for predicting functional language and MD network regions based on anatomical alignment and task-free data.
- Examined mechanisms of how high-level language functions are established by intra-hemispheric connections between the temporal and frontal cortices (lesion case study).
- Wrote a Python toolbox for natural language processing, including part-of-speech tagging, lemmatization, lexical feature mapping and sentence-level representations. The aim was to generate lexical and semantic word embedding spaces.
- Exploited Representational Similarity Analysis (RSA) to investigate how activation patterns in language and MD functional regions of interest can be explained by lexical, syntactic and semantic features.
First-authored paper in prep: Tuckute, G., Mineroff, Z., Blank, I., Kean, H., Fedorenko, E. (2020): Intra-Hemispheric Temporal-Frontal Connections Appear Necessary to Wire Up Left Frontal Cortex for Language, Cortex.
GitHub toolbox: Natural language processing and lexical feature analyses.
Semantic decoding of visual stimuli using EEG
- Designed an experiment for EEG-based semantic decoding of naturalistic, visual stimuli and acquired EEG data from healthy subjects in non-laboratory settings.
- Processed EEG signals and applied various feature engineering approaches for model optimization: wavelet decomposition, temporal binning, power signals.
- Implemented data dimensionality reduction methods: t-distributed stochastic neighbor embedding (t-SNE), principal component analysis (PCA) and multidimensional scaling (MDS).
- Created decoding classification models in EEG: Support Vector Machines (SVM), dense neural networks (NN), convolutional neural networks (CNN) in a leave-one-subject-out approach.
- Extracted representational image feature vectors from experimental stimuli using an Inception-V3 net and created regression models to map high-dimensional EEG feature vectors to image feature vectors (ridge regression and dense NNs).
- Performed Representational Similarity Analysis (RSA) of EEG across subjects and evaluated inter-subject correlation using bootstrapping and permutation testing.
- Proposed a method for computing and evaluating sensitivity maps for EEG-based SVM classificiation models with Radial Basis Function (RBF) kernels, and implemented an NPAIRS cross-validation framework for estimation of effect size uncertainty.
GitHub toolbox: SVM decoding and sensitivity mapping.
