Explore an advanced tutorial on unsupervised discovery of temporal sequences in high-dimensional datasets, focusing on a novel tool called seqNMF. Delve into the challenges of identifying interpretable, low-dimensional features in large-scale neural recordings and learn how seqNMF extends convolutional non-negative matrix factorization techniques to extract significant sequences from neural data. Discover the tool's application to various neural and behavioral datasets, and gain hands-on experience with demo code. Understand the cross-validation procedures for assessing factor significance, choosing optimal parameters, and validating results on held-out data. Engage with practical examples, simulations, and real-world applications in neuroscience, equipping yourself with cutting-edge techniques for analyzing complex temporal patterns in high-dimensional data.
Unsupervised Discovery of Temporal Sequences in High-Dimensional Datasets
Overview
Syllabus
Intro
Neurons form sequences
Non-negative matrix factorization (NMF)
In practice ... redundant factors
Simple multiplicative update rules
Testing seqNMF on simulated sequences
SeqNMF factorizations are highly consistent
Testing significance of each factor on held-out data
Method to choose lambda
Cross-validation procedure for NMF
Cross-validation procedure for convolutional NMF
Taught by
MITCBMM
