Cerebellar learning using perturbations

Guy Bouvier, Johnatan Aljadeff, Claudia Clopath, Célian Bimbard, Jonas Ranft, Antonin Blot, Jean‐Pierre Nadal, Nicolas Brunel, Vincent Hakim, Boris BarboureLife 2018 · DOI

TL;DR

This paper proposes a novel computational model for how the cerebellum learns complex movements, suggesting it uses a form of "stochastic gradient descent with estimated global errors" (SGDEGE). This framework explains how the brain can assign credit for errors to specific neural pathways, overcoming limitations of current theories, and is supported by new experimental findings under more physiological conditions.

What Problem Does This Paper Address?

The cerebellum is crucial for learning and executing fast, coordinated movements. The prevailing theory suggests that "complex spikes" (a type of Purkinje cell firing) signal movement errors, leading to the depression of active "parallel fiber" synapses. However, this theory struggles with the "credit assignment problem": how to translate a single, global error signal into specific adjustments for thousands of individual synapses across many neurons and movements. This is particularly challenging for complex, arbitrary movements where the error-to-correction mapping isn't fixed. The authors aim to bridge this gap by proposing a biologically plausible algorithm that can solve this credit assignment problem.

Methods

The authors propose a computational framework based on stochastic gradient descent with estimated global errors (SGDEGE). Key elements include:

  • Perturbations: Spontaneous "complex spikes" (triggered by climbing fibers) act as perturbations, subtly altering movement and creating "eligibility traces" at active synapses.
  • Error Change Extraction: A novel mechanism is proposed where the inferior olive adaptively cancels the average movement error. This allows for the extraction of the change in error (ΔE) that results from the perturbation. This cancellation is hypothesized to occur via plasticity at mossy fiber-to-nucleo-olivary neuron synapses.
  • Plasticity Rules: Based on this framework, the authors predict specific plasticity rules for parallel fiber synapses. These rules differ from the current consensus, predicting potentiation (LTP) under certain conditions where depression (LTD) is expected.
  • Experimental Validation: To test these predictions, they conducted in vitro electrophysiology experiments on mouse cerebellar slices. Crucially, they used conditions designed to be more "physiological" (e.g., lower extracellular calcium, more dispersed parallel fiber stimulation) than typical plasticity studies.
  • Simulations: They developed reduced computational models and simulations to analyze the convergence properties and storage capacity of the SGDEGE algorithm, comparing it to simpler models and the "delta rule."

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • Complex Spikes as Perturbations: The paper reframes the role of complex spikes from purely error signals to active perturbations that probe the motor system, enabling learning.
  • Error Cancellation for Credit Assignment: The proposed mechanism of adaptively canceling the average error is a clever way to isolate the change in error caused by a specific perturbation, which is essential for solving the credit assignment problem.
  • Physiological Conditions Matter: The experimental results strongly suggest that standard in vitro plasticity protocols might not reflect in vivo learning due to differences in experimental conditions, particularly calcium levels and stimulation patterns.
  • Reinterpreting Plasticity: The predicted plasticity rules, which seem counterintuitive compared to the consensus, are supported by experiments under more physiological conditions, suggesting a need to re-evaluate existing findings.
  • Broader Applicability: The authors propose that similar SGDEGE principles might also apply to learning in other brain regions, such as the basal ganglia.

Key Results

  • Experimental Support for Novel Plasticity Rules: The experiments provided evidence supporting the predicted plasticity rules, including LTP when granule cell and climbing fiber activity were simultaneous (GP_), and LTD when an error complex spike followed (GPE_). These contrasted with control conditions (G__ and G_E) where less plasticity was observed.
  • Algorithm Convergence: Simulations demonstrated that the SGDEGE algorithm can converge to optimize motor outputs, even for multiple patterns learned in parallel.
  • Storage Capacity: The theoretical analysis suggests the SGDEGE algorithm has a storage capacity comparable to that of an optimal analog perceptron.
  • Learning Speed: While SGDEGE is slower than learning algorithms that use full error information (like the delta rule), it is argued to be more biologically plausible for complex movements where full error information is unavailable.

Limitations and Future Work

  • Experimental Conditions: While the authors used more physiological conditions, they acknowledge that replicating in vivo conditions perfectly is challenging. The precise mechanisms by which these altered conditions affect plasticity are not fully detailed.
  • Error Cancellation Site: The exact cellular locus of the error cancellation mechanism (e.g., mossy fiber-to-nucleo-olivary neuron synapses) is hypothesized and requires further investigation.
  • Movement Specificity: The model assumes that error cancellation is movement-specific, which is plausible but needs further experimental validation.
  • "Bicycle Problem": The paper discusses the "bicycle problem" (how to use diverse error signals for complex movements) and proposes "output convergence" and "error broadcast" solutions, which are areas for future research.
  • Basal Ganglia Application: The suggestion that SGDEGE might operate in the basal ganglia is speculative and requires dedicated investigation.
  • Learning Rate Adaptation: The authors mention extensions to SGDEGE that incorporate adaptive learning rates, which could accelerate learning and are a promising avenue for future work.

Summary updated

Abstract

The cerebellum aids the learning of fast, coordinated movements. According to current consensus, erroneously active parallel fibre synapses are depressed by complex spikes signalling movement errors. However, this theory cannot solve the <i>credit assignment problem</i> of processing a global movement evaluation into multiple cell-specific error signals. We identify a possible implementation of an algorithm solving this problem, whereby spontaneous complex spikes perturb ongoing movements, create eligibility traces and signal error changes guiding plasticity. Error changes are extracted by adaptively cancelling the average error. This framework, <i>stochastic gradient descent with estimated global errors</i> (SGDEGE), predicts synaptic plasticity rules that apparently contradict the current consensus but were supported by plasticity experiments in slices from mice under conditions designed to be physiological, highlighting the sensitivity of plasticity studies to experimental conditions. We analyse the algorithm's convergence and capacity. Finally, we suggest SGDEGE may also operate in the basal ganglia.
BibTeX
@inproceedings{bouvier2018cerebellar,
  author = {Guy Bouvier and Johnatan Aljadeff and Claudia Clopath and Célian Bimbard and Jonas Ranft and Antonin Blot and Jean‐Pierre Nadal and Nicolas Brunel and Vincent Hakim and Boris Barbour},
  title = {{Cerebellar learning using perturbations}},
  booktitle = {eLife},
  year = {2018},
  doi = {10.7554/elife.31599},
}

Loading related content...

Figure