Lights in the Brain

(Advanced) In this video, I present the main ideas and results that my thesis work during the past half year has led to. This is a preliminary video for a later video explaining the results in more laymen terms. Abstract Brain-computer interfaces (BCI) provide a new framework for studying and augmenting the human mind and have been successful in offering new perspectives for what human cognition is capable of. Implanted BCI enable patients to control robot arms and write using just their thoughts, however, BCIs on the head (non-invasive BCIs) are limited in how easily and reliably they can be implemented in everyday use because of algorithm limitations. With recent revolutions in artificial language processing, we can imagine novel solutions to these algorithm problems. I attempt to combine these fields and present the practical implementation of solutions to make the use of neuroimaging devices for everyday use better. I perform self-supervised training (LeCun & Misra, 2021) to pre-train a machine learning model using the LSTM architecture (Hochreiter & Schmidhuber, 1997) on functional near-infrared spectroscopic (fNIRS) neuroimaging data (Naseer & Hong, 2015) from the NIRx NIRSport2 system (NIRx, 2021) and transfer and fine-tune it for a BCI thought classification task (Yoo et al., 2018) as is done with language models (C. Sun et al., 2019). As far as I am aware, this is the first example of such work. 🔗 Resources - Thesis Github: https://github.com/esbenkc/fnirs-bci - My website: https://kran.ai - Playlist for this series: https://www.youtube.com/watch?v=oLulfdNI3E0&list=PLQYlYK0hzM-AGXLej1J8SIiLtFKx79zCX&ab_channel=EsbenKran 📜 My thesis paper - On Github:https://github.com/esbenkc/fnirs-bci/blob/main/Lights%20in%20the%20Brain.pdf 📚 Topics Covered - Reviewing and summarising my machine learning and neuroimaging thesis work - Transfer learning with LSTM on brain data - Understanding and transferring understanding of natural language processing (NLP) theory to functional near-infrared spectroscopy (fNIRS) neuroimaging