Abstract
The ability to identify interpretable, low-dimensional features that capture the dynamics of large-scale neural recordings is a major challenge in neuroscience. Repeated temporal patterns (sequences) are not succinctly captured by traditional dimensionality reduction techniques, so neural data is often aligned to behavioral task references. We describe a task-independent, unsupervised method, which we call seqNMF, that provides a framework for extracting sequences from high-dimensional datasets, and assessing the significance in held-out data. We test seqNMF on simulated datasets under a variety of noise conditions, and also on several neural datasets. In a hippocampal dataset, seqNMF identifies neural sequences that match those calculated manually by reference to behavioral events. In a songbird dataset, seqNMF discovers abnormal motor sequences in birds that lack stereotyped songs. Thus, by identifying temporal structure directly from neural data, seqNMF enables dissection of complex neural circuits in the absence of reliable temporal references from stimuli or behavioral outputs.