Poster F108, Tuesday, March 28, 8:00 – 10:00 am, Pacific Concourse
Deep learning techniques for decoding EEG signatures of viewing or refreshing face, scene, and word stimuli
Jacob Williams1, Ashok Samal1, Matthew Johnson1; 1University of Nebraska - Lincoln
Modern deep learning techniques have proven revolutionary in the classification of images, speech signals, and other data types, yet are rarely applied in the analysis of cognitive neuroscience datasets. Deep learning techniques are particularly effective in data with high temporal and spatial correlation, seemingly making them a natural fit for techniques such as fMRI and EEG. However, due to the tendency for deep learning techniques to overfit, thousands or even millions of samples are usually required for optimal results. Cognitive neuroscience datasets rarely reach that size; thus multivariate pattern analysis (MVPA) in neuroimaging typically employs simpler classification techniques, which conversely may ignore relevant features and underfit. We explored whether deep learning techniques could be fruitfully applied to EEG MVPA in a dataset previously analyzed with Sparse Multinomial Logistic Regression (SMLR). Participants in the study viewed and refreshed (thought back to) face, scene, or word pictures, and we classified the category either perceived or refreshed. We found that in individual-subject models (~200 samples per participant) analogous to typical MVPA approaches, deep learning techniques only matched SMLR after extensive regularization, data augmentation, and careful tuning. Furthermore, complicated architectures using convolutional or recurrent layers drastically overfitted even in the presence of regularization and data augmentation. However, when considering a universal (cross-subject) model containing several thousand samples, convolutional and recurrent architectures significantly outperformed both individual-subject and universal SMLR models for visual perception. These results suggest that deep learning techniques may significantly boost MVPA performance in neuroimaging studies when sufficient data are available.
Topic Area: METHODS: Neuroimaging