The contemporary methods in Machine Learning, called Deep Learning, owe its success to two major factors: An exponential increase of computational power and availability of abundant data. Problems with computational power is solved with the availability of parallel processsing units and cloud computing. Unfortunately data problem still continues to be an issue. In some special fields, such as, brain decoding it is very difficult and time consuming to acquire data, which is statistically sufficient to train a deep neural network. An alternative to generate data is to reuse the available "statistically similar data", so that the information of the available data can be transferred to relatively small datasets.
Brain decoding involves analyzing the cognitive states of human brain by using some statistical techniques to understand the relationships among the cognitive states, based on Neuroimaging data. A very powerful tool to acquire the brain data is functional magnetic resonance images (fMRI), which generates three dimensional brain volume at each time instant, while a subject performs a cognitive task, such as , emotion, memory, social activities etc.
Recently, instead of using low level techniques to study the brain activities, machine learning methods gain some popularity to investigate a large variety of properties of brain and its functions.
In this thesis, we investigate the relationships among well-defined cognitive tasks, based on fMRI data. The suggested approach is expected to be useful for transferring the learning capabilities across the cognitive tasks from the abundant data to relatively scarce data.
We propose a pipeline to create and verify a graph, which shows relations between the states of a brain during several cognitive tasks. This pipeline consists of three steps:
In the first step, we create the graph using an existing end-to-end method and applying to an fMRI dataset. In this step we find a latent representation for every task in the dataset and compare them to create an affinity matrix. However, due to scarcity of data we need to verify this structure. In the second step, we suggest a simple method to quantify the stability of the graph by comparing different parameters. In the last step, we use the graph adjacency matrix for transferring the information across the cognitive tasks.