Direct neural perturbations reveal a dynamical mechanism for robust computation

Daniel J. O’Shea, Lea Duncker, Werapong Goo, Xulu Sun, Saurabh Vyas, Eric M. Trautmann, Ilka Diester, Charu Ramakrishnan, Karl Deisseroth, Maneesh Sahani, Krishna V. ShenoybioRxiv 2022 · DOI

TL;DR

This paper uses direct neural stimulation in primate motor cortex to uncover the underlying dynamical system governing reaching movements. They demonstrate that motor cortex activity is organized within a low-dimensional, self-contained subspace, which is crucial for robust computation and is causally linked to behavior.

What Problem Does This Paper Address?

Mammalian behavior relies on complex neural circuits that generate robust and flexible patterns of activity. While associative studies have linked specific neural activity patterns to behaviors, they haven't fully explained the dynamical mechanisms shaping these patterns. Specifically, the authors investigate whether neural circuits act as high-dimensional "reservoirs" capable of generating many patterns, or if they are constrained to low-dimensional subspaces tailored for specific tasks. Understanding this dynamical structure is key to deciphering how neural circuits compute and generate complex behaviors.

Methods

The researchers employed a combination of experimental and computational techniques:

  • Optogenetic and Electrical Microstimulation (ICMS): They directly stimulated primate motor cortex during reaching tasks to perturb neural activity. Optogenetics provided a broad, additive perturbation, while ICMS offered a more localized and potentially distorting perturbation.
  • Latent Variable Modeling (LVM): They developed and utilized a novel analytical approach to model neural population activity. This involved fitting latent linear dynamical systems (LDS) and a more complex E/I network model to capture the underlying dynamics.
  • ERAASR2 Algorithm: A new artifact removal technique was developed to handle the significant electrical interference from ICMS, allowing for accurate neural recordings during stimulation.
  • Schur Decomposition: Used to analyze the structure of the dynamical system and understand the flow of information between different modes of activity.

Datasets

The study used neural recordings from two macaques (Q and O) performing a reaching task. Data was collected during both optogenetic and ICMS stimulation experiments. No new benchmarks were introduced; rather, existing datasets were analyzed and modeled.

Key Insights and Ideas

  • Motor Cortex Operates in a Low-Dimensional Subspace: Contrary to a high-dimensional reservoir model, motor cortical activity during reaching is confined to a specific, low-dimensional subspace. This subspace is "self-contained," meaning its dynamics are largely independent of activity outside of it.
  • Task Dynamics Subspace is Robust: This task-relevant subspace is oriented in a way that makes it robust to strong "non-normal amplification," a phenomenon where small perturbations can be transiently amplified. This robustness is crucial for reliable computation.
  • Causal Link Between Neural Subspace and Behavior: Perturbations in motor cortex only affected reaching kinematics when they altered neural states within this specific task dynamics subspace. This provides strong evidence for a causal role of this subspace in controlling movement.
  • E/I Network Model Explains Dynamics: A balanced excitatory-inhibitory (E/I) network model, constrained by Dale's Law and E/I balance, successfully reproduced the observed dynamics and stimulation responses, suggesting this biological constraint is key to generating the observed low-dimensional, self-contained structure.
  • ICMS Distorts, Optogenetics Translates: ICMS induced a "contraction" of neural states within the task dynamics space, effectively distorting its geometry and leading to behavioral effects. Optogenetic stimulation, in contrast, primarily caused an additive translation, largely preserving the subspace's geometry.

Key Results

  • Optogenetic stimulation resulted in additive neural responses that decayed rapidly and did not alter reach kinematics, ruling out high-dimensional reservoir dynamics.
  • ICMS produced more complex effects, including a contraction of neural states within the task dynamics space, which was correlated with changes in hand velocity.
  • An E/I network model, fit to non-stimulated data, could accurately predict the responses to both optogenetic stimulation and ICMS, supporting the idea of subspace-structured dynamics.
  • The task dynamics space was found to be aligned with the E/I co-activation pattern, a consequence of Dale's Law, which contributes to both robustness and the observed stimulation responses.

Limitations and Future Work

  • Optogenetic Stimulation as a Special Case: The authors note that purely additive effects from optogenetics might be a special case, and other stimulation modalities (inhibition, strong activation) could have different effects.
  • Distinguishing H2 and H3: While the study strongly supports subspace-structured dynamics (H2), distinguishing it from path-following dynamics (H3) was challenging with optogenetics alone. ICMS provided stronger evidence against H3.
  • Generalizability: The study focused on motor cortex and reaching. Future work could explore whether these principles of self-contained, low-dimensional subspaces apply to other brain areas and cognitive functions.
  • Role of Non-Normality: The paper highlights the role of non-normal dynamics in E/I networks for robust computation, but further investigation into how this mechanism is leveraged across different circuits is warranted.

Summary updated

Abstract

Abstract The rich repertoire of skilled mammalian behavior is the product of neural circuits that generate robust and flexible patterns of activity distributed across populations of neurons. Decades of associative studies have linked many behaviors to specific patterns of population activity, but association alone cannot reveal the dynamical mechanisms that shape those patterns. Are local neural circuits high-dimensional dynamical reservoirs able to generate arbitrary superpositions of patterns with appropriate excitation? Or might circuit dynamics be shaped in response to behavioral context so as to generate only the low-dimensional patterns needed for the task at hand? Here, we address these questions within primate motor cortex by delivering optogenetic and electrical microstimulation perturbations during reaching behavior. We develop a novel analytic approach that relates measured activity to theoretically tractable, dynamical models of excitatory and inhibitory neurons. This computational model captures the dynamical effects of these perturbations and demonstrates that motor cortical activity during reaching is shaped by a self-contained, low-dimensional dynamical system. The subspace containing task-relevant dynamics proves to be oriented so as to be robust to strong non-normal amplification within cortical circuits. This task dynamics space exhibits a privileged causal relationship with behavior, in that stimulation in motor cortex perturb reach kinematics only to the extent that it alters neural states within this subspace. Our results resolve long-standing questions about the dynamical structure of cortical activity associated with movement, and illuminate the dynamical perturbation experiments needed to understand how neural circuits throughout the brain generate complex behavior.
BibTeX
@misc{oshea2022direct,
  author = {Daniel J. O’Shea and Lea Duncker and Werapong Goo and Xulu Sun and Saurabh Vyas and Eric M. Trautmann and Ilka Diester and Charu Ramakrishnan and Karl Deisseroth and Maneesh Sahani and Krishna V. Shenoy},
  title = {{Direct neural perturbations reveal a dynamical mechanism for robust computation}},
  year = {2022},
  doi = {10.1101/2022.12.16.520768},
  howpublished = {bioRxiv},
}
Figure