Invariant neural dynamics drive commands to control different movements

Vivek R. Athalye, Preeya Khanna, Suraj Gowda, Amy L. Orsborn, Rui M. Costa, Jose M. CarmenabioRxiv 2021 · DOI

TL;DR

This study reveals that the brain uses "invariant neural dynamics" – consistent patterns of neural activity transitions – to generate a wide range of movements. By using a brain-machine interface (BMI), researchers found that these dynamics are not just descriptive of behavior but actively drive motor commands, allowing for flexible and efficient control.

What Problem Does This Paper Address?

The nervous system's remarkable ability to produce a vast repertoire of movements is thought to stem from reusing underlying neural codes rather than having separate neural circuits for each movement. Previous research suggested that the dynamics of neural population activity (how activity patterns change over time) are similar across different movements. However, it remained unclear whether these invariant dynamics are actively used by the brain to issue motor commands, or if they are merely a byproduct of movement. This paper directly tests this by investigating the causal link between invariant neural dynamics and the commands that control movement.

Methods

The core of this research involved a brain-machine interface (BMI) in rhesus macaques. The BMI translated neural activity from the motor cortex into commands for a prosthetic cursor. This allowed researchers to precisely measure and manipulate the transformation from neural activity to motor commands.

Key methodological aspects include:

  • Brain-Machine Interface (BMI): Used to decode motor cortex activity into cursor movements, providing a known and controlled transformation from neural signals to commands.
  • Invariant Dynamics Modeling: The researchers modeled neural activity transitions using linear dynamical systems ($x_{t+1} = A x_t + b + \text{input}_t + \text{noise}_t$), where the matrix $A$ captures the invariant dynamics.
  • Optimal Feedback Control (OFC) Model: A hierarchical model was developed to understand how the brain might integrate invariant dynamics with feedback to generate commands efficiently.
  • Behavior-Preserving Shuffle Analysis: A crucial control method where neural activity was shuffled while preserving the temporal sequence of commands. This allowed researchers to distinguish between dynamics that are truly predictive of commands versus those that are merely correlated with behavior.

Datasets

The study used neural recordings from two adult male rhesus macaques (Monkey G and Monkey J) performing two neuroprosthetic tasks: a center-out task and an obstacle-avoidance task. These tasks generated a variety of movement conditions. The data included binned spike counts, cursor trajectories, condition parameters, and decoder parameters. This data has been deposited in the DANDI Archive.

Key Insights and Ideas

  • Invariant Dynamics as Command Drivers: The most significant finding is that invariant neural dynamics are not just descriptive but actively drive motor commands. This means the brain uses these consistent patterns of neural activity transitions to generate different movements.
  • Flexible Command Generation: The same motor command can be issued by different neural activity patterns, but these patterns transition according to the same invariant dynamics. This allows for flexibility in how a command is generated.
  • Efficiency Through Dynamics: The OFC model suggests that invariant dynamics help the brain control movement more efficiently by reducing the amount of input needed, especially when integrating feedback. This implies that the brain leverages these intrinsic dynamics to minimize effort.
  • Decoding Null Space Importance: The study highlights that invariant dynamics operate significantly in the "decoder null space" – dimensions of neural activity that don't directly influence the BMI command. This suggests that these dynamics might be preparing for future commands or handling internal processing without immediately affecting the output.

Key Results

  • Same Command, Different Patterns: The study demonstrated that the same BMI command could be issued by distinct neural activity patterns across different movement conditions.
  • Predictable Transitions: These different neural activity patterns transitioned according to low-dimensional, invariant dynamics, predicting the next command.
  • Low-Dimensional Dynamics: The identified invariant dynamics were low-dimensional (around 4 dimensions), suggesting a compact representation of movement control.
  • OFC Model Efficiency: Simulations showed that the OFC model incorporating invariant dynamics required significantly less input to control movement compared to a model without these dynamics. This effect was dependent on the alignment of dynamics with the decoder.
  • Generalization: The invariant dynamics models generalized well to new commands and conditions not seen during training.

Limitations and Future Work

  • Model Simplification: While the study used linear dynamics models, the authors acknowledge that real neural systems are non-linear. However, they found that non-linear models did not significantly outperform the linear ones in their experiments.
  • BMI vs. Natural Movement: The study used a BMI, and while it provides a controlled environment, the authors speculate that the observed decaying dynamics might differ from rotational dynamics seen in natural arm movements, potentially due to the BMI's reliance on feedback control.
  • Future Directions: The authors suggest future research should explore:
  • The specific neural connectivity that gives rise to invariant dynamics.
  • How the brain sends inputs to leverage these dynamics.
  • Whether invariant dynamics directly drive muscles during physical movement.
  • The role of invariant dynamics in learning new behaviors and motor rehabilitation.

Summary updated

Abstract

Summary It has been proposed that the nervous system has the capacity to generate a wide variety of movements because it re-uses some invariant code. Previous work has identified that dynamics of neural population activity are similar during different movements, where dynamics refer to how the instantaneous spatial pattern of population activity changes in time. Here we test whether invariant dynamics of neural populations are actually used to issue the commands that direct movement. Using a brain-machine interface that transformed rhesus macaques’ motor cortex activity into commands for a neuroprosthetic cursor, we discovered that the same command is issued with different neural activity patterns in different movements. However, these different patterns were predictable, as we found that the transitions between activity patterns are governed by the same dynamics across movements. These invariant dynamics are low-dimensional, and critically, they align with the brain-machine interface, so that they predict the specific component of neural activity that actually issues the next command. We introduce a model of optimal feedback control that shows that invariant dynamics can help transform movement feedback into commands, reducing the input that the neural population needs to control movement. Altogether our results demonstrate that invariant dynamics drive commands to control a variety of movements, and show how feedback can be integrated with invariant dynamics to issue generalizable commands.
BibTeX
@misc{athalye2021invariant,
  author = {Vivek R. Athalye and Preeya Khanna and Suraj Gowda and Amy L. Orsborn and Rui M. Costa and Jose M. Carmena},
  title = {{Invariant neural dynamics drive commands to control different movements}},
  year = {2021},
  doi = {10.1101/2021.08.27.457931},
  howpublished = {bioRxiv},
}
Figure