Abstract
Cerebellar computations are necessary for fine behavioural control and are thought to rely on internal probabilistic models performing state estimation. We propose that the cerebellum infers how states contextualise (i.e., interact with) each other, and coordinates extra-cerebellar neuronal dynamics underpinning a range of behaviours. To support this claim, we describe a cerebellar model for state estimation that includes states interactions, and link the underlying inference with the neuronal architecture and dynamics observed empirically. This is formalised using the free energy principle, which provides a dual perspective on a system in terms of both the dynamics of its physical – in this case neuronal – states, and the inference process they entail. As a proof of principle, we simulate cerebellar-dependent synchronisation of whisking and respiration, which are known to be tightly coupled in rodents. In summary, we suggest that cerebellar-dependent contextualisation of behaviour can explain its ubiquitous involvement in most aspects of behaviour.
Introduction
The cerebellum presents a significant anatomical enigma: it contains more neurons than the rest of the brain combined, yet its internal structure is remarkably simple and repetitive. Despite this uniform architecture, it is involved in a vast array of functions, from fine motor control and language to social interaction and decision-making. The central challenge in cerebellar research is explaining how one uniform architecture supports such functional diversity. While traditionally viewed as a "forward model" predicting sensory consequences, this paper argues for a broader role: the cerebellum acts as a high-level "contextualizer" that coordinates extra-cerebellar dynamics in the cortex and brainstem to ensure behavioral harmony.
Theoretical Framework and Methods
The authors utilize the Free Energy Principle (FEP) and Active Inference to model the cerebellum as an inference engine. In this framework, biological systems minimize "surprise" (prediction error) through internal probabilistic models. The proposed model suggests a division of labor: extra-cerebellar structures handle the "content" of behavior, while the cerebellum provides the "context"—learned expectations of how different states should interact.
By acting as a functional hierarchy above other systems, the cerebellum provides "empirical priors" that guide lower-level dynamics. This allows for resilient state estimation even when individual systems are subjected to noise or external disruptions.
Biological Implementation
A key strength of this model is its direct mapping onto cerebellar microcircuitry. The mathematical variables of the inference engine correspond to specific neuronal types and pathways observed empirically.
In this mapping:
- Mossy fibers (mf) provide the observations ($y$).
- Granule cells (grc) represent state and velocity estimates ($\mu_x, \mu_{x'}$).
- Purkinje cells (Pjc) encode the "hidden causes" ($\mu_v$) that coordinate behavior.
- Glomeruli (glom) and Cerebellar Nuclei (CN) facilitate the calculation of prediction errors ($\epsilon_y$).
Results: Resilient Behavioral Coupling
To validate the model, the authors simulated two non-linear Kuramoto oscillators representing whisking and respiration—two systems that are tightly coupled in rodents but have no direct physical connection in the simulation.
Key findings include:
- Resilient Synchrony: The cerebellar model successfully forced the two systems into tight synchronization even when they had different intrinsic speeds. Without this contextualization, the behaviors drifted apart.
- Noise and Perturbation Tolerance: The model maintained coordination despite significant Gaussian noise. When a physical perturbation was introduced (e.g., an abrupt phase change in whisking), the cerebellar state estimator rapidly recognized the mismatch and re-synchronized the systems.
- Ataxia Simulation: Removing the cerebellar output from the model replicated the loss of coordination seen in cerebellar ataxia, where individual movements remain possible but their integration and timing fail.
Discussion and Future Work
The uniform architecture of the cerebellum appears optimized for linear state estimation applicable to any behavioral domain. However, the current model relies on linear approximations, and real cerebellar processing likely involves non-linearities. Furthermore, while the model addresses inference, it does not yet explicitly simulate the learning dynamics associated with the olivary-climbing fiber system. Future research will focus on scaling this "contextual glue" to coordinate more complex, multi-stream sensory and motor dynamics.
Summary updated