Unsupervised discovery of temporal sequences in high-dimensional datasets, with applications to neuroscience.

Emily L Mackevicius,Michale S Fee,Mark S Goldman,Shijie Gu,Natalia I Denisenko,Andrew H Bahle,Alex H Williams

doi:10.7554/elife.38471

Abstract

Identifying low-dimensional features that describe large-scale neural recordings is a major challenge in neuroscience. Repeated temporal patterns (sequences) are thought to be a salient feature of neural dynamics, but are not succinctly captured by traditional dimensionality reduction techniques. Here, we describe a software toolbox-called seqNMF-with new methods for extracting informative, non-redundant, sequences from high-dimensional neural data, testing the significance of these extracted patterns, and assessing the prevalence of sequential structure in data. We test these methods on simulated data under multiple noise conditions, and on several real neural and behavioral datas. In hippocampal data, seqNMF identifies neural sequences that match those calculated manually by reference to behavioral events. In songbird data, seqNMF discovers neural sequences in untutored birds that lack stereotyped songs. Thus, by identifying temporal structure directly from neural data, seqNMF enables dissection of complex neural circuits without relying on temporal references from stimuli or behavioral outputs.

Highlights

The ability to detect and analyze temporal sequences embedded in a complex sensory stream is an essential cognitive function, and as such is a necessary capability of neuronal circuits in the brain (Clegg et al, 1998; Janata and Grafton, 2003; Bapi et al, 2005; Hawkins and Ahmad, 2016), as well as artificial intelligence systems (Cui et al, 2016; Sutskever et al, 2014)
Recent advances in technology for electrophysiological and optical measurements of neural activity have enabled the simultaneous recording of hundreds or thousands of neurons (Chen et al, 2013; Kim et al, 2016; Scholvin et al, 2016; Jun et al, 2017), in which neuronal dynamics are often structured in sparse sequences (Hahnloser et al, 2002; Harvey et al, 2012; MacDonald et al, 2011; Okubo et al, 2015; Fujisawa et al, 2008; Pastalkova et al, 2008)
We evaluated the performance of convolutional nonnegative matrix factorization (convNMF) with and without the x-ortho penalty on datasets with a larger number of sequences

Summary

Introduction

The ability to detect and analyze temporal sequences embedded in a complex sensory stream is an essential cognitive function, and as such is a necessary capability of neuronal circuits in the brain (Clegg et al, 1998; Janata and Grafton, 2003; Bapi et al, 2005; Hawkins and Ahmad, 2016), as well as artificial intelligence systems (Cui et al, 2016; Sutskever et al, 2014). Recent advances in technology for electrophysiological and optical measurements of neural activity have enabled the simultaneous recording of hundreds or thousands of neurons (Chen et al, 2013; Kim et al, 2016; Scholvin et al, 2016; Jun et al, 2017), in which neuronal dynamics are often structured in sparse sequences (Hahnloser et al, 2002; Harvey et al, 2012; MacDonald et al, 2011; Okubo et al, 2015; Fujisawa et al, 2008; Pastalkova et al, 2008) Such sequences can be identified by averaging across multiple trials, but only in cases where an animal receives a temporally precise sensory stimulus, or executes a sufficiently stereotyped behavioral task.

Methods

Results

Conclusion