Sparse atomic feature learning via gradient regularization: With applications to finding sparse representations of fMRI activity patterns

We present an algorithm, Sparse Atomic Feature Learning (SAFL), that transforms noisy labeled datasets into a sparse domain by learning atomic features of the underlying signal space via gradient minimization. The sparse signal representations are highly compressed and cleaner than the original signals. We demonstrate the effectiveness of our techniques on fMRI activity patterns. We produce low-dimensional, sparse representations which achieve over 98% compression of the original signals. The transformed signals can be used to classify left-out testing data at a higher accuracy than the initial data.