Template-aligned Transfer Learning on Brain Decoding Problem.

Brain decoding involves a set of methods to estimate the brain activities that correspond to the brain signals, acquired via fMRI or similar techniques. Acquisition of fMRI data is a costly and hard process. Due to this, it is important to utilize readily available fMRI data. In this study, we develop an incremental method that can do transfer learning between the available source fMRI datasets and target fMRI datasets. This method produces aligned features by calculating the generalizable relation between cognitive tasks. Also, it uses a matrix transformation that minimizes the mismatch between subjects and experimentation processes belonging to the same cognitive task. Aligned features obtained from different datasets transfer knowledge with a transfer learning algorithm and decodes the cognitive tasks on the target fMRI dataset. We observed that, the template based solution has shown on average a 18% performance increase compared to our baseline model.