Abstract
Animals rapidly adapt their movements to external perturbations, a process paralleled by changes in neural activity in the motor cortex. Experimental studies suggest that these changes originate from altered inputs ($H_{input}$) rather than from changes in local connectivity ($H_{local}$), as neural covariance is largely preserved during adaptation. Since measuring synaptic changes in vivo remains very challenging, we used a modular recurrent neural network to qualitatively test this interpretation. As expected, $H_{input}$ resulted in small activity changes and largely preserved covariance. Surprisingly, given the presumed dependence of stable covariance on preserved circuit connectivity, $H_{local}$ led to only slightly larger changes in activity and covariance, still within the range of experimental recordings. This similarity is due to $H_{local}$ only requiring small, correlated connectivity changes for successful adaptation. Simulations of tasks that impose increasingly larger behavioural changes revealed a growing difference between $H_{input}$ and $H_{local}$, which could be exploited when designing future experiments.
Introduction
When learning a new motor skill or adjusting to a change in visual feedback—such as using a computer mouse with a different sensitivity—the brain adapts within minutes. This rapid motor learning is reflected in the shifting firing patterns of neurons in the primary motor cortex (M1) and dorsal premotor cortex (PMd). A fundamental question in motor neuroscience is whether these changes are driven by "upstream" inputs (e.g., from the cerebellum) or by "local" plasticity within the motor cortex itself.
Historically, the stability of neural covariance—the statistical relationship between the activities of different neurons—has been used as evidence against local plasticity. The logic was that if the physical "wiring" of the circuit changed, the statistical structure of its activity should also change. Consequently, many researchers concluded that local connections are fixed during short-term adaptation. This study uses modular recurrent neural networks (RNNs) to challenge this conclusion, demonstrating that small, highly coordinated synaptic changes can drive learning while remaining virtually invisible to standard covariance analysis.
Modeling the Motor Hierarchy
To investigate these competing hypotheses, the authors employed a modular RNN architecture designed to mimic the hierarchy of the motor system. The model consists of three distinct modules:
- Upstream Module: Represents areas providing input to the motor cortex.
- PMd Module: Simulates the dorsal premotor cortex.
- M1 Module: Simulates the primary motor cortex, which generates the final motor output.
The network was initially trained to perform a standard "center-out" reaching task. To simulate adaptation, the researchers introduced a visuomotor rotation (VR), where visual feedback is rotated by a specific angle (e.g., 30°), requiring the model to adjust its output to reach the target.
The researchers compared two scenarios: $H_{input}$, where only the connections from the upstream module were allowed to change, and $H_{local}$, where only the internal connections within the PMd and M1 modules were plastic.
Results: The "Invisible" Plasticity of Local Circuits
The most striking finding of the study is that for small perturbations (like a 30° rotation), the $H_{local}$ model produces neural signatures that are nearly identical to those of the $H_{input}$ model. Both models successfully preserved neural covariance, matching experimental data recorded from macaques.
The reason for this "invisible" plasticity lies in the magnitude and coordination of the synaptic changes. In the $H_{local}$ model, successful adaptation required an average weight change of only 1-2%. These changes were not random; they were "low-dimensional" and highly correlated across the population. Because the changes were so coordinated, they did not disrupt the existing covariance structure, explaining why previous experimental studies might have overlooked local plasticity.
The study also explored "stress tests" to differentiate the two hypotheses. For larger perturbations, such as 90° rotations or "reassociation" tasks (where the mapping between visual cues and reach directions is shuffled), the models predicted diverging neural signatures. For instance, in 90° rotations, $H_{input}$ predicts larger activity changes in PMd than M1, whereas $H_{local}$ predicts the opposite.
Discussion and Future Directions
This research demonstrates that stable neural patterns do not necessarily imply stable physical connections. The motor cortex may be significantly more plastic during short-term learning than previously assumed. The finding that small, correlated changes can drive adaptation provides a new perspective on how the brain balances flexibility with stability.
While the current RNN model is a simplification—lacking real-time feedback loops like proprioception—it provides a powerful framework for generating testable predictions. The authors suggest that future experiments using "reassociation" tasks will be crucial for definitively identifying the locus of rapid motor learning. Such tasks would force the system into a regime where the differences between upstream and local plasticity become experimentally detectable.
Summary updated