Unsupervised Discovery of Demixed, Low-Dimensional Neural Dynamics across Multiple Timescales through Tensor Component Analysis

Abstract

Perceptions, thoughts, and actions unfold over millisecond timescales, while learned behaviors can require many days to mature. While recent experimental advances enable large-scale and long-term neural recordings with high temporal fidelity, it remains a formidable challenge to extract unbiased and interpretable descriptions of how rapid single-trial circuit dynamics change slowly over many trials to mediate learning. We demonstrate a simple tensor component analysis (TCA) can meet this challenge by extracting three interconnected, low-dimensional descriptions of neural data: neuron factors, reflecting cell assemblies; temporal factors, reflecting rapid circuit dynamics mediating perceptions, thoughts, and actions within each trial; and trial factors, describing both long-term learning and trial-to-trial changes in cognitive state. We demonstrate the broad applicability of TCA by revealing insights into diverse datasets derived from artificial neural networks, large-scale calcium imaging of rodent prefrontal cortex during maze navigation, and multielectrode recordings of macaque motor cortex during brain machine interface learning.

Sequential and efficient neural-population coding of complex task information

“…In other words, the nominally high-dimensional neural population activity of all surveyed cortical areas can be parsimoniously described by a lowdimensional set of multiplicative factors. This type of factorizable neural responses seems intriguingly ubiquitous, as similar findings have been reported for mouse prefrontal cortex and nonhuman-primate motor cortex 83 . As each neuron has a characteristic timepreference (Fig.…”

Thalamic activity patterns unfolding over multiple time scales predict seizure onset in absence epilepsy

“…We built on existing tensor-decomposition and time-warp models by developing a method that seeks to find a low-dimensional representation of neural dynamics over time and trials, which simultaneously allows for variable temporal warping of the dynamics across trials. The result is a mixture of the well studied canonical polyadic decomposition (CPD) (also known as tensor-components analysis or TCA), a method designed to represent high-dimensional data tensors as a summation of low-dimensional factors 55 , and time-warping, designed to allow for temporal warping of neural activity across trials to uncover meaningful neural representations and reduce noise introduced by temporal jitter. 66 Our motivation for combining these models was two fold: to improve the robustness of TCA to temporal jitter, and also to formulate a unified statistical model to identify trial-specific amplitude and temporal modulation of neural activity, both of which have been shown to emerge in real neural recordings.…”

“…66 Our motivation for combining these models was two fold: to improve the robustness of TCA to temporal jitter, and also to formulate a unified statistical model to identify trial-specific amplitude and temporal modulation of neural activity, both of which have been shown to emerge in real neural recordings. 55,66 We point the interested reader to our manuscript describing the method in more detail 67 .…”

“…Action potential firing from hundreds of neurons is now reliably obtained with multi-electrode arrays positioned in multiple brain regions. These studies commonly utilize matrix decomposition techniques to digest highdimensional neural data (action potential times from multiple neural "units") and extract low-dimensional representations of the population activity (often termed "latent" or "hidden" population dynamics) [55][56][57][58] . These population dynamics capture much of the variability observed in the raw data in a much more compressed form, while simultaneously providing an interpretable representation of how the neural activity evolves in response to external or internal cues.…”

Unsupervised Clusterless Decoding using a Switching Poisson Hidden Markov Model

“…A complimentary paradigm to the supervised decoding approaches mentioned above, is that of unsupervised approaches to understand and interpret neural data (see e.g., Low et al, 2018;Williams et al, 2018;Maboudi et al, 2018;Mackevicius et al, 2019). These unsupervised approaches provide powerful ways of understanding the latent dynamics of neural activity, but to date they all require sorted spikes from ensembles of neurons.…”