The notion that the fundamental unit of cortical computation may not be the individual neuron but rather the neuronal population is gaining ground as recording techniques make it possible to capture the activity of ever-larger groups of cells. Chronically implanted electrode arrays offer the possibility of tracking this activity over several weeks, or even several months — a prerequisite for studying long-term learning mechanisms. However, exploiting these data raises two challenges: their high dimensionality on the one hand, and signal drift over time on the other, whether it stems from electrode displacement or degradation or from task-induced plasticity. Hidden Markov models (HMMs) are a promising tool for describing this activity in terms of discrete latent states, but previous approaches did not account for the particular statistical properties of neuronal spiking data, did not accommodate longitudinal data, and did not model condition-specific differences.
To overcome these limitations, the authors propose a multilevel Bayesian HMM built on three innovations: multivariate log-normal Poisson emission distributions, multilevel parameter estimation, and the introduction of trial-specific condition covariates. This framework was applied to multi-unit spiking data recorded in the primary motor cortex of macaques performing a cued reach-to-grasp-and-place task. The model was trained solely from the spike trains, without any information about the timing of behavioral events.
Despite this "blind" learning, the model identifies population states closely tied to behavioral events: movement onset, object contact, and placement. The statistics of these states also match the durations of the inter-event intervals. In both animals, an initial state reflects a static phase of low activity, whereas a later state coincides with movement initiation; other states correspond to the reaching and placing phases. Model comparison selects six states for one monkey and five for the other, a difference the authors attribute to more ballistic movements and a smaller number of trials in the second animal. Importantly, the association between states and behaviors remains consistent across several days of recording, a consistency that a single-level HMM fails to reproduce from one session to the next.
The authors highlight several limitations, notably the assumptions inherent to HMMs (only one active state at a time, dependence solely on the current state), the use of multi-unit rather than single-unit activity, and a slight underestimation of the proportion of zero spike counts. Demonstrated here on an already-learned task, this approach is, in their view, particularly well suited to studying the long-term plasticity of neuronal populations in the context of motor learning.