Unsupervised Discovery of Temporal Sequences in Neural Data
Alexander Williams, Stanford University
The ability to identify interpretable, low-dimensional features that capture the dynamics of large-scale neural recordings is a major challenge in neuroscience. Traditional dimensionality reduction techniques do not succinctly capture repeated temporal patterns (sequences), so neural data is often aligned to behavioral task references. We describe a method based on convolutive nonnegative matrix factorization that provides a framework for extracting sequences from high-dimensional data sets, and on assessing the significance in held-out data. We test this method on simulated data sets under a variety of noise conditions and also on several neural data sets. In a hippocampus data set, the model identifies neural sequences that match those calculated manually by reference to behavioral events. In a songbird data set, the model discovers abnormal motor sequences in birds that lack stereotyped songs. Thus, by identifying temporal structure directly from neural data, the methodology we develop enables dissection of complex neural circuits in the absence of reliable temporal references from stimuli or behavioral outputs.
Abstract Author(s): Alex H. Williams, Emily L. Mackevicius, Andrew H. Bahle, Shijie Gu, Natalia I. Denissenko, Mark S. Goldman, Michael S. Fee