Project Significance:
By hyperalignment and regression model, we will be able to create an optimal decoder that can decide whether target Subject X is seeing target Object X, or something else. The weight vector k will give us the optimal combination and demonstrates how different subjects’ data for each category, as well as for RS-connectivity, may help to form the best decoding model for SX. Once we have a model for predicting Subject X’s BOLD responses for voxels for some OX as opposed to other random objects, the decoding for application of fear reduction in clinical settings should follow easily.
By applying machine learning techniques to fMRI data, we can train an RNN to create a generative model that matches the ongoing spontaneous fluctuation of brain activities. Hidden Markov models are great tools to isolate noisy signals in the RS-connectivity and allow us to model only for the activation of voxel patterns that are relevant, i.e. the input nodes in RNN. Once we have a model for RS-connectivity, we do not have to blindly hyperalign the entire voxel pattern of the brain. Instead, we can focus on the feedback in activation levels of the nodes in the RNN, and align these networks among brains. These models will standardize the fear reduction process introduced in Koizumi et al.
Machine learning has not been employed to analyze voxel patterns due to the relatively small size of fMRI data. Hyperalignment allows us to study data across brains and makes deep learning (LeCun, Bengio, & Hinton, 2015) applicable. We can find better representations, instead of vector spaces, to model the high-level abstractions in voxel patterns and create deep network models to learn such data.
