Find Brain Regions Critical For Language Learning from Neuroimaging with Transformer-Based Model

Animal experiments prevalent in neuroscience pharmacology face challenges such as small sample size, complicated experimental design, and limited generalization to human beings. Non-invasive neuroimaging approach can help uncover the interplay between different regions of the brain with fewer compliance restrictions, paving the path to decode the interconnectivity between different regions of the brain. Deep learning models are especially promising when applied to neuroimaging to circumvent the limitations with traditional approach. Although small sample size remains to be problematic, transformers, the building blocks of large-language models, can be a good fit for tackling these challenges. With the recent trend of applying transformers to computer vision domain such as the famous ViT [3], this study attempts to address these challenges by applying transformer-based models such as NiT and MINiT proposed by the study [1] to the “ADHD-200” [2] dataset. The resulting attention map discovers more brain regions related to strong language learning performance.