Neuroscience meets Deep Learning

We attempt to classify the cognitive thought process of human subjects based on their brain activity observed through functional Magnetic Resonance Imaging(fMRI) using convolutional neural networks. This project has a huge potential in clinical and health applications. The viability of this project has been shown in previous related work. The primary goal of moving this approach forward is to gauge if, with reasonable probability, it is possible to train classifiers across many subjects. The main reason of using CNNs is that most other classifiers require feature engineering and pre-processing. With the use of neural networks, we avoid this and attempt to show, via comparison with a baseline SVM classifier (which has feature extraction done), that CNNs can perform better even with raw data input. We show our different approaches to model these classifiers and report our results which encapsulate the degree of success achieved over 9 different subjects’ fMRI data.