TL;DR
This study investigates how the brain selects among redundant neural activity patterns that produce the same motor output. Using a brain-computer interface (BCI) in monkeys, researchers found that redundant neural activity isn't freely chosen but is coupled to task-relevant activity, suggesting inherent constraints on how the brain can leverage redundancy for computation.
What Problem Does This Paper Address?
The brain often uses multiple, seemingly equivalent neural activity patterns to achieve the same motor outcome – a phenomenon known as neural redundancy. This redundancy is thought to be beneficial for computations like movement preparation or stable processing. However, it's unclear what principles govern the selection of these redundant patterns. Do they minimize energy, are they largely unconstrained, or are there other factors at play? This paper aims to uncover these constraints by precisely controlling and observing redundant neural activity during a BCI task.
Methods
The researchers trained three Rhesus monkeys to control a cursor on a screen using neural activity from their primary motor cortex via a brain-computer interface (BCI). This BCI setup allowed them to precisely define which neural activity patterns were "redundant" – meaning they produced the same cursor movement. They then developed several hypotheses for how the brain might select among these redundant patterns, drawing inspiration from motor control theories (e.g., minimal energy, minimal intervention). These hypotheses' predictions about the distribution of "output-null" neural activity (activity that doesn't affect the cursor's movement) were compared against the monkeys' actual neural activity. The core of their approach involved:
- BCI Control: Monkeys learned to map neural activity to cursor movement.
- Defining Redundancy: Using the known BCI mapping, they could identify output-potent (cursor-moving) and output-null dimensions of neural activity.
- Hypothesis Testing: They formulated and tested hypotheses about how the brain chooses output-null activity, comparing predicted distributions to observed ones.
- Internal Model Estimation: They accounted for the possibility that the monkey's internal model of the BCI mapping might differ from the actual mapping.
Datasets
The study used neural recordings from three Rhesus macaques performing a 2D center-out BCI task. Data from these experiments were previously analyzed in related studies (Sadtler et al., 2014; Golub et al., 2018). The authors specifically selected sessions where animals demonstrated stable control of the BCI, using a rigorous criterion for learning and consistency.
Key Insights and Ideas
- Muscular Redundancy Principles Don't Directly Apply: Theories that explain how the body minimizes muscle effort (like minimal firing or minimal deviation) failed to accurately predict the observed patterns of redundant neural activity. This suggests neural redundancy is governed by different principles than muscular redundancy.
- Coupling of Redundant and Task-Relevant Activity: The most surprising finding was that the distribution of redundant (output-null) neural activity was strongly coupled to the task-relevant (output-potent) activity. This means the brain doesn't freely choose any redundant pattern; the choice is constrained by the activity needed for the task.
- "Fixed Distribution" Hypothesis Wins: The "Fixed Distribution" hypothesis, which posits that the relationship between output-potent and output-null activity remains consistent across different BCI mappings, best explained the observed data. This implies that neural activity patterns are "yoked" across dimensions.
- Limits on Computational Exploitation: This coupling suggests that the brain's ability to use redundant neural activity for computations beyond motor control (e.g., planning, information processing) might be limited by the demands of the primary motor task.
Key Results
- Hypotheses based on minimal energy principles (Minimal Firing, Minimal Deviation) showed significant errors in predicting output-null activity (average histogram error of 73% and 31%, respectively).
- "Uncontrolled" hypotheses, suggesting output-null activity is largely unconstrained, performed better but still fell short (Uncontrolled-uniform: 57% error; Uncontrolled-empirical: 24% error).
- The "Fixed Distribution" hypothesis, which assumes a coupled relationship between output-potent and output-null activity, provided the best predictions, with an average histogram error of only 13.4%.
- Analysis of variance changes in dimensions that transitioned from output-potent to output-null also supported the Fixed Distribution hypothesis, showing minimal changes in variance, contrary to predictions from minimal energy models.
Limitations and Future Work
- Timescale: The study focused on short timescales (hours to days) of learning. It's possible that over longer periods, neural redundancy constraints might change. Future work could investigate this with extended training.
- Dimensionality: The current BCI task had a relatively low-dimensional output (2D cursor movement). Future research could explore higher-dimensional effectors (e.g., robotic limbs) to see if similar coupling principles hold.
- OFC and Cost Functions: While the study didn't directly support standard Optimal Feedback Control (OFC) models, the authors suggest that designing specific cost functions or applying OFC to different levels of the motor system might yield predictions consistent with their findings.
- Understanding the "Why": While the study identifies that output-null activity is coupled, the precise neural mechanisms driving this coupling remain an open question. Future work could explore how these dependencies impact computations beyond motor control.
Summary updated