A stable, long - term cortical signature underlying consistent behavior

Juan Álvaro Gallego, Matthew G. Perich, Raeed H. Chowdhury, Sara A. Solla, Lee E. Miller

TL;DR

This paper reveals that the underlying "latent dynamics" of neural populations in sensorimotor cortex remain remarkably stable over years, even as individual neurons change. This stability in neural dynamics, rather than individual neurons, is the key to consistent learned behavior and allows for robust decoding of movement over long periods.

What Problem Does This Paper Address?

We all perform learned tasks with impressive consistency over time, from riding a bike to playing a musical instrument. Intuitively, we assume the brain mechanisms supporting these behaviors must also be stable. However, directly observing this neural stability has been challenging. Traditional neural recordings often show significant changes in the activity of individual neurons over days or weeks due to neuron turnover and other factors. This makes it difficult to determine if the underlying neural computations are stable or if the brain is constantly recalibrating. This paper tackles this by investigating whether the low-dimensional dynamics of neural populations, which represent the coordinated activity patterns, are stable over long periods, and if this stability explains consistent behavior.

Methods

The core of this work lies in a clever approach to track "latent dynamics" despite changing neural recordings.

  • Neural Manifold and Latent Dynamics: The authors model neural activity as lying on a low-dimensional "manifold" within a high-dimensional neural space. The "latent dynamics" are the time-varying activity along the principal axes (neural modes) of this manifold. Principal Component Analysis (PCA) is used to identify these modes and estimate the latent dynamics.
  • Alignment of Latent Dynamics: The key innovation is a method to "align" latent dynamics recorded on different days. Since the empirical neural manifold (based on recorded neurons) changes, the estimated latent dynamics can appear different even if the underlying true dynamics are stable. Canonical Correlation Analysis (CCA) is used to find linear transformations that maximize the correlation between latent dynamics from different days, effectively compensating for the changing neural recordings.
  • Decoding and Prediction: Linear models (Wiener filters for kinematics, naive Bayes classifiers for target direction) were trained on neural activity or aligned latent dynamics from one day and tested on data from subsequent days to assess the stability of predictive power.

Datasets

The study utilized neural and behavioral data collected from six monkeys performing a center-out reaching task.

  • Monkeys: Six monkeys (C, M, T, J, H, P) were trained.
  • Task: An instructed-delay reaching task using a planar manipulandum.
  • Recording Duration: Data was collected over periods ranging from ~20 to ~750 days, allowing for long-term stability analysis.
  • Cortical Areas: Recordings were made from dorsal premotor cortex (PMd), primary motor cortex (M1), and primary somatosensory cortex (S1).
  • Data Type: Multiunit activity (threshold crossings) was primarily used, with some analyses also employing sorted single neurons.

Key Insights and Ideas

  • Latent Dynamics are Stable, Neurons are Not: The most striking finding is that while individual neurons show significant turnover and changes in their tuning properties over time, the underlying low-dimensional latent dynamics of the neural population remain remarkably stable for years.
  • Alignment is Key: The CCA-based alignment method is crucial. Without it, latent dynamics appear to change significantly between days. With alignment, they become highly similar, even across months and years.
  • Stability Underlies Consistent Behavior: The stable latent dynamics are directly linked to consistent behavioral performance. Models trained on these aligned dynamics can predict behavior accurately even after long delays, far better than models based on raw neural activity.
  • Cortical Origin of Stability: The stability is not simply a byproduct of motor mechanics or peripheral feedback. Stable dynamics were observed in PMd even during the planning phase before movement execution.
  • A Fundamental Building Block: The authors propose that these stable latent dynamics are not just a convenient mathematical abstraction but are the fundamental, stable building blocks that the brain uses to generate consistent, learned behaviors.

Key Results

  • Latent Dynamics Stability: Aligned latent dynamics in M1, PMd, and S1 remained stable for up to two years, with normalized similarity values close to 1 (meaning they were almost as correlated as dynamics within the same day).
  • Neuron Turnover: The percentage of matched neurons across days dropped to below 30% after 25 days, highlighting the dynamic nature of individual neural recordings.
  • Predictive Accuracy: Linear models trained on aligned latent dynamics predicted hand kinematics with high accuracy (normalized predictive accuracy of ~0.96) even when tested on data from days far apart (up to two years). In contrast, models based on raw neural activity degraded significantly over time.
  • PMd Planning: Classification accuracy for predicting reach targets from PMd planning activity using aligned latent dynamics was nearly as good across days as within-day models, while raw neural activity models performed poorly across days.

Limitations and Future Work

  • Single Task Focus: The study focused on a single, stereotyped reaching behavior. The authors acknowledge that different behaviors would likely require their own distinct latent dynamics. Future work could explore how latent dynamics are shared or differ across multiple tasks.
  • Linear Transformations: While CCA effectively aligns dynamics, the authors discuss the theoretical possibility that it might be "too powerful" and find linear transformations even if true dynamics have fundamentally changed. However, they argue against this based on previous work and the consistency of their findings.
  • Brain Area Specificity: The stability might decrease in brain areas further removed from direct motor output or sensory input, as these areas integrate more complex internal states. Future research could investigate this gradient of stability across the cortex.
  • BCI Applications: The findings have significant implications for brain-computer interfaces (BCIs), suggesting that using aligned latent dynamics could lead to more stable and long-lasting control, reducing the need for frequent recalibration.

Summary updated

Abstract

Animals readily execute learned motor behaviors in a consistent manner over long periods of time, yet similarly stable neural correlates remained elusive up to now. How does the cortex achieve this stable control? Using the sensorimotor system as a model of cortical processing, we investigated the hypothesis that the dynamics of neural latent activity, which capture the dominant co-variation patterns within the neural population, are preserved across time. We recorded from populations of neurons in premotor, primary motor, and somatosensory cortices for up to two years as monkeys performed a reaching task. Intriguingly, despite steady turnover in the recorded neurons, the low-dimensional latent dynamics remained stable. Such stability allowed reliable decoding of behavioral features for the entire timespan, while fixed decoders based on the recorded neural activity degraded substantially. We posit that latent cortical dynamics within the manifold are the fundamental and stable building blocks underlying consistent behavioral execution.[ES]
BibTeX
@inproceedings{gallego2018stable,
  author = {Juan Álvaro Gallego and Matthew G. Perich and Raeed H. Chowdhury and Sara A. Solla and Lee E. Miller},
  title = {{A stable, long - term cortical signature underlying consistent behavior}},
  booktitle = {DIGITAL.CSIC},
  year = {2018},
  doi = {10.20350/digitalcsic/8573},
}

Loading related content...

Figure