Sparse Autoencoders for Word Decoding from Magnetoencephalography

An open problem in the processing and analysis of neuroimaging data is finding the optimal set of features from which to decode or study underlying neural processes. Handcrafted features sets have been examined in the past with varying degrees of success. Here, we explore sparse autoencoders, a method to extract high-level feature representations of neural data in an unsupervised way, without any manual feature engineering. We pair this unsupervised feature creation step with a L2 regularized regression, which is robust to overfitting. We apply this method to the task of single word decoding from Magnetoencephalography (MEG) data which is known for its high degree of noise and complexity. We show that, for some subjects, sparse autoencoding is advantageous, though it does not produce an overall increase in decoding accuracy. Our results represent the first step towards a neural-network system that could join feature extraction and decoding into one powerful joint system.