Brain Decoding Using fNIRS

Authors: Lu Cao, Dandan Huang, Yue Zhang, Xiaowei Jiang, Yanan Chen12602-12611

AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We investigate brain decoding tasks under the help of f NIRS and empirically compare f NIRS with f MRI. Primarily, we find that: 1) like f MRI scans, activation patterns recorded from f NIRS encode rich information for discriminating concepts, but show limits on the possibility of decoding fine-grained semantic clues; 2) f NIRS decoding shows robustness across different brain regions, semantic categories and even subjects; 3) f NIRS has higher accuracy being decoded based on multi-channel patterns as compared to single-channel ones, which is in line with our intuition of the working mechanism of human brain. Our findings prove that f NIRS has the potential to promote a deep integration of NLP and cognitive neuroscience from the perspective of language understanding. We release the largest f NIRS dataset by far to facilitate future research. ... To this end, we conduct one pilot study and one large-scale decoding experiment using f NIRS technology.
Researcher Affiliation Academia Lu Cao1 , Dandan Huang2,3 , Yue Zhang2,3 , Xiaowei Jiang4, Yanan Chen4 1Singapore University of Technology and Design, Singapore 2School of Engineering, Westlake University, China, 3 Institute of Advanced Technology, Westlake Institute for Advanced Study, China 4Henan University, China
Pseudocode No The paper describes the methods textually but does not include any explicit pseudocode or algorithm blocks.
Open Source Code Yes We release the largest f NIRS dataset by far to facilitate future research. ... 1https://github.com/caolusg/decoding fnirs ... we release the largest f NIRS dataset by far for future research, which covers 50 objects from 10 semantic categories.
Open Datasets Yes We release the largest f NIRS dataset by far to facilitate future research, which covers 50 objects from 10 semantic categories. 1https://github.com/caolusg/decoding fnirs
Dataset Splits Yes The linear regression model is trained and tested by leave-two-out and leave-one-out cross validation. In the leave-two-out approach, the model is trained repeatedly using C(N 2 N ) stimuli and tested using the two stimuli left out. In the leave-one-out approach, the model is trained using C(N 1 N ) stimuli and tested using the one left out. The procedure repeats until all stimuli have been trained and tested.
Hardware Specification No Participants blood oxygen levels are measured by the NIRx NIRScout f NIRS system throughout the exposure. (This specifies the fNIRS device, not the computational hardware used for data processing or model training). No further specific hardware for computation is mentioned.
Software Dependencies No Preprocessing of f NIRS data is performed using nirs Lab (Xu, Graber, and Barbour 2014). ... The word2vec (Mikolov et al. 2013) and Glo Ve embedding (Pennington, Socher, and Manning 2014) were superior to others. We adopt the Glo Ve embedding... we use the ridge regression to learn a linear map... (No version numbers are given for nirs Lab, word2vec, GloVe, or any other software or libraries.)
Experiment Setup No For each subject and his/her stimulitriggered Hb O variation, we use the ridge regression to learn a linear map w : Hj V by minimizing the function: J = ||w Hj V ||2 2 + α||w||2, (2) where α is a regularization hyperparameter. ... we test the influence of word embedding dimension size on the decoding performance, with Glo VE embedding sizes of 50, 100, 200 and 300. ... Hence, we fix the dimension of word vector as 50 in the following analysis. ... We test model performance under various decoding time window sizes. ... Hence, we fix the decoding time window as 6.5-7.5 seconds in the following experiments. (While some parameters are explored and then fixed (embedding dim, time window), the critical `alpha` hyperparameter value for ridge regression is not specified.)