Improving Brain Computer Interfaces Using Deep Scale-Invariant Temporal History Applied to Scalp Electroencephalogram Data

Brain Computer Interface (BCI) applications employ machine learning to decode neural signals through time to generate actions. One issue facing such machine learning algorithms is how much of the past they need to decode the present. DeepSITH (Deep Scale-Invariant Temporal History), is a deep neural network with layers inspired by how the mammalian brain represents recent vs. less-recent experience. A single SITH layer maintains a log-compressed representation of the past that becomes less accurate with older events, unlike other approaches that maintain a perfect copy of events regardless of how far in the past they occurred. By stacking layers of this compressed representation, we hypothesized that DeepSITH would be able to decode patterns of neural activity from farther in the past and combine them efficiently to guide the BCI in the present. We tested our approach with the Kaggle "Grasp and Lift challenge" dataset. This motor movement dataset has 12 subjects, 10 series of 30 grasp and lift trials per subject, with 6 classes of events to decode. We benchmark DeepSITH performances on this dataset against another common machine learning technique for integrating features over extended time scales, long short-term memory (LSTM). DeepSITH reproducibly achieves higher accuracy in predicting motor movement events than LSTM, and also takes significantly fewer epochs and less memory to train, in comparison to LSTM. In summary, DeepSITH can efficiently process more data, with increased prediction accuracy and learning speed. This result shows that DeepSITH is an advantageous model to consider when developing BCI technologies.