Internal models for interpreting neural population activity during sensorimotor control

Matthew D. Golub, Byron M. Yu, Steven M. ChaseeLife 2015 · DOI

TL;DR

This paper introduces a novel framework (IME) to extract "internal models" – the brain's predictive understanding of how its motor commands affect the body – from neural activity in monkeys controlling a brain-machine interface (BMI). They found that mismatches between these internal models and the actual BMI system explain a significant portion of movement errors and speed control issues, suggesting internal models are crucial for efficient motor control.

What Problem Does This Paper Address?

The brain is remarkably adept at controlling our bodies, even with sensory feedback delays. A leading hypothesis is that the brain uses "internal models" to predict the consequences of motor commands before sensory feedback arrives. However, directly observing and quantifying these internal models in real-time has been challenging. This paper tackles the problem of how to extract these internal models from high-dimensional neural activity and, crucially, to determine if and how mismatches between these internal models and the actual physical system (in this case, a BMI) contribute to errors in motor control.

Methods

The authors developed an Internal Model Estimation (IME) framework. This statistical tool uses recorded neural activity (from primary motor cortex) and cursor feedback in a closed-loop BMI task to infer:

  1. The subject's internal model: A mathematical representation of the monkey's beliefs about how its neural commands translate into cursor movement. This model operates on a timescale of tens of milliseconds.
  2. Moment-by-moment internal state predictions: These are visualized as "whiskers" showing the predicted cursor trajectory based on the internal model.

Key aspects of the IME framework include:

  • Forward Model: It assumes the brain predicts future states based on current inputs and its internal model.
  • Probabilistic Model: It uses expectation-maximization (EM) to estimate model parameters.
  • Cross-validation: Rigorous validation ensures the models aren't just overfitting the data.
  • Focus on High-Dimensional Neural Activity: Unlike previous approaches, IME leverages the rich information in population-level neural recordings.

Datasets

The study used data collected from two rhesus macaques performing a 2D brain-machine interface (BMI) task. The monkeys were trained to move a cursor to targets using neural signals. The data included:

  • Neural recordings from primary motor cortex.
  • Cursor position and velocity data.
  • Target locations.

The paper does not introduce new benchmark datasets but rather analyzes data collected specifically for this experiment, including conditions with an "intuitive" BMI mapping and a "perturbed" BMI mapping to study adaptation.

Key Insights and Ideas

  • Internal Models are Dynamic and Mismatched: The brain doesn't just passively reflect the external world; it actively constructs predictive models. Crucially, these models aren't always perfectly aligned with reality, and this "mismatch" is a significant source of error.
  • Explaining Errors Beyond Noise: Traditional explanations for motor errors often focus on random neural noise. This work demonstrates that structured errors, explainable by internal model mismatches, are a major contributor.
  • High-Dimensional Neural Activity is Key: The power of IME comes from analyzing the collective activity of many neurons, revealing patterns that low-dimensional analyses (like just looking at cursor velocity) would miss.
  • Internal Models Drive Adaptation: When the BMI mapping was changed (perturbed), the monkeys' internal models adapted to better match the new reality, showcasing the brain's ability to learn and update its predictions.
  • Speed Control Deficiencies Explained: Long-standing issues with controlling cursor speed in BMIs can be attributed to the internal model's systematic over/underestimation of intended speeds compared to the actual BMI output.

Key Results

  • 65% of Movement Errors Explained: Mismatches between the subjects' internal models and the actual BMI mapping explained approximately 65% of cursor movement errors.
  • Compensation for Sensory Delays: Monkeys demonstrated an ability to compensate for visual feedback delays by internally predicting cursor position, suggesting they were using an internal model.
  • Speed Control Mismatch: Internal models systematically overestimated low intended speeds and underestimated high intended speeds, explaining deficiencies in BMI speed control.
  • Adaptation to Perturbations: When the BMI mapping was perturbed, the extracted internal models dynamically adjusted to better match the new mapping, showing adaptive learning.

Limitations and Future Work

  • Focus on M1: The study primarily analyzes neural activity from the primary motor cortex (M1). While M1 is downstream of internal model computations, the exact neural circuits implementing these models (e.g., cerebellum, parietal cortex) are not directly investigated.
  • Simplified Internal Model: The IME framework uses a forward model and assumes a "straight-to-target" aiming intention. More complex objectives (e.g., minimizing jerk) could potentially improve explanatory power.
  • Learning Limitations: The study suggests that the speed of learning and adaptation might be a bottleneck in BMI performance, especially when the internal model is already "good enough" and there's less pressure to adapt. Future work could explore how to design BMI mappings that better leverage or accelerate this adaptation process.
  • Improving BMI Design: The authors suggest that incorporating insights from internal models could lead to improved BMI designs, potentially by creating mappings that better align with the user's internal predictions.

Summary updated

Abstract

To successfully guide limb movements, the brain takes in sensory information about the limb, internally tracks the state of the limb, and produces appropriate motor commands. It is widely believed that this process uses an internal model, which describes our prior beliefs about how the limb responds to motor commands. Here, we leveraged a brain-machine interface (BMI) paradigm in rhesus monkeys and novel statistical analyses of neural population activity to gain insight into moment-by-moment internal model computations. We discovered that a mismatch between subjects' internal models and the actual BMI explains roughly 65% of movement errors, as well as long-standing deficiencies in BMI speed control. We then used the internal models to characterize how the neural population activity changes during BMI learning. More broadly, this work provides an approach for interpreting neural population activity in the context of how prior beliefs guide the transformation of sensory input to motor output.
BibTeX
@inproceedings{golub2015internal,
  author = {Matthew D. Golub and Byron M. Yu and Steven M. Chase},
  title = {{Internal models for interpreting neural population activity during sensorimotor control}},
  booktitle = {eLife},
  year = {2015},
  doi = {10.7554/elife.10015},
}
Figure