Multiple tasks viewed from the neural manifold: Stable control of varied behavior

Juan Álvaro Gallego, Matthew G. Perich, Stephanie Naufel, Christian Éthier, Sara A. Solla, Lee E. MillerbioRxiv 2017 · DOI

TL;DR

This paper proposes that the brain's motor cortex controls a wide range of behaviors not by having neurons encode specific movements, but by flexibly combining a set of stable "neural modes." These modes, reflecting underlying network connectivity, are remarkably consistent across different tasks, suggesting a fundamental building block for motor control.

What Problem Does This Paper Address?

How does the brain, specifically the primary motor cortex (M1), achieve the incredible flexibility to perform a vast array of movements? Traditionally, researchers looked for specific neurons that "encoded" particular aspects of movement (like direction or speed). However, this approach often led to contradictory findings, as individual neurons seemed to change their "coding" depending on the task. This paper tackles this puzzle by proposing an alternative: that M1's flexibility arises from the population-level activity of neurons, organized into stable "neural modes" that can be combined in different ways to generate diverse behaviors.

Methods

The authors employed several key techniques:

  • Neural Recordings: High-density microelectrode arrays were used to record the activity of hundreds of neurons simultaneously in the hand area of M1 in monkeys.
  • Principal Component Analysis (PCA): Used to identify "neural modes" – principal components that capture the dominant patterns of neural activity and span a low-dimensional "neural manifold." This manifold represents the subspace of neural activity that is actually explored during behavior.
  • Principal Angles: A geometric measure to quantify the similarity in orientation between neural manifolds derived from different tasks.
  • Canonical Correlation Analysis (CCA): Used to compare the temporal dynamics (how the neural modes are activated over time) of these neural modes across different tasks.
  • Demixed PCA (dPCA): A more advanced technique to disentangle neural activity patterns related to specific behavioral parameters (time, task, target location) from task-independent components.
  • Linear Decoders: Used to predict muscle activity (EMG) from neural activity, testing how well neural modes could account for motor commands.

Datasets

The study utilized datasets from three rhesus macaque monkeys performing a variety of upper limb motor tasks. These included:

  • Wrist Tasks: One-dimensional isometric, movement, and elastic-loaded movement tasks, as well as a two-dimensional isometric task.
  • Reach-to-Grasp Tasks: A "ball" task (reaching, grasping, transporting, dropping) and a "power grip" task (reaching, grasping, squeezing).

These tasks were designed to elicit distinct patterns of neural and muscle activity.

Key Insights and Ideas

  • Stable Building Blocks: The core idea is that M1 doesn't reinvent the wheel for every movement. Instead, it uses a set of "neural modes" that are like fundamental building blocks, shaped by the brain's fixed connectivity.
  • Population-Level Organization: The similarities in neural modes are only apparent when looking at the entire population of neurons, not individual cells. This highlights the importance of collective neural activity.
  • Task-Independent Components: A significant portion of neural activity is organized into modes that are not specific to any particular task or target, suggesting these modes represent more general computational processes.
  • Flexible Combination is Key: The brain's ability to generate diverse behaviors comes from its capacity to flexibly combine these stable neural modes in different ways.
  • Predictive Power for Muscle Commands: Specific, task-independent neural modes can accurately predict muscle activity, suggesting a direct link between these fundamental neural patterns and motor output.

Key Results

  • Similar Neural Manifolds: Neural manifolds (the space explored by neural activity) for different tasks showed remarkably similar orientations, indicating that the underlying structure of neural covariation is preserved.
  • Preserved Neural Mode Dynamics: The temporal activation patterns of several neural modes were also highly correlated across different tasks, even when the overall neural and muscle activity patterns were very different.
  • Task-Independent Mapping to Muscles: A subset of neural modes, particularly those related to target location, provided a consistent, task-independent mapping to muscle commands (EMG). Decoders using these modes predicted EMG activity with high accuracy (R² of 0.52 ± 0.32), outperforming decoders based on other types of modes.
  • dPCA Reveals Behavioral Parameters: Demixed PCA successfully identified neural modes that specifically tracked time, target location, task, or combinations thereof, further dissecting the functional organization of neural activity.

Limitations and Future Work

  • Stereotypical Tasks: The study used laboratory-defined tasks. The authors suggest that studying more complex, natural behaviors would be a valuable extension.
  • Dimensionality Limits: The dimensionality of M1 manifolds appears to be relatively low for these tasks. However, theoretical work suggests dimensionality might increase with task complexity, potentially requiring more recorded neurons than currently feasible to fully map.
  • Generalizability: While evidence suggests similar mechanisms might apply to other cortical areas and non-motor functions, further research is needed to confirm this.
  • Optogenetic/Stimulation Studies: The authors note that while evidence for network connectivity shaping neural modes is strong, direct experimental manipulation (e.g., optogenetics) could provide even more definitive proof.

Summary updated

Abstract

Abstract How do populations of cortical neurons have the flexibility to perform different functions? We investigated this question in primary motor cortex (M1), where populations of neurons are able to generate a rich repertoire of motor behaviors. We recorded neural activity while monkeys performed a variety of wrist and reach-to-grasp motor tasks, each requiring a different pattern of neural activity. We characterized the flexibility of M1 movement control by comparing the “neural modes” that capture covariation across neurons, believed to arise from network connectivity. We found large similarities in the structure of the neural modes across tasks, as well as striking similarities in their temporal activation dynamics. These similarities were only apparent at the population level. Moreover, a subset of these well-preserved modes captured a task-independent mapping onto muscle commands. We hypothesize that this system of flexibly combined, stable neural modes gives M1 the flexibility to generate our wide-ranging behavioral repertoire.
BibTeX
@misc{gallego2017multiple,
  author = {Juan Álvaro Gallego and Matthew G. Perich and Stephanie Naufel and Christian Éthier and Sara A. Solla and Lee E. Miller},
  title = {{Multiple tasks viewed from the neural manifold: Stable control of varied behavior}},
  year = {2017},
  doi = {10.1101/176081},
  howpublished = {bioRxiv},
}
Figure