Fast and slow synaptic plasticity enables concurrent control and learning

Reviewer #1 (Public review):

Summary:

This paper proposes a new set of local synaptic plasticity rules that differs from classic rules in two regards: First, working under the assumption that signals coming into synapses change smoothly over time and thus have temporal correlations such that immediate activity is positively correlated with subsequent activity, it proposes both fast plasticity that immediately corrects errors as well as slower plasticity. Second, it derives these rules from optimal, Bayesian control theory principles that, even without the fast component of plasticity, are shown to provide more accurate performance than classic, non-Bayesian plasticity rules. As a proof of principle, it applies these to a simple cerebellar learning example that demonstrates how the proposed rules lead to learning performance that exceeds that achieved with classic cerebellar learning rules. The work also provides a potential normative explanation for post-climbing fiber spike pauses in Purkinje cell firing and proposes testable predictions for cerebellar experiments. Overall, I found the idea to be compelling and potentially broadly applicable across many systems. Further, I thought the work was a rare, very beautiful display of the application of optimal control theory to fundamental problems in neuroscience. My comments are all relatively minor and more expressions of interest than criticism.

Comments:

(1) The algorithm assumes, reasonably, that inputs are relatively smooth. However, I was wondering if this could make additional experimental predictions for the system being exceptionally noisy or otherwise behaving in signature ways if one were able to train a real biological network to match a rapidly changing or non-smooth function that does not align with the underlying assumptions of the model.

(2) The algorithm assumes that one can, to a good approximation, replace individual input rates by their across-synapse average. How sensitive is the learning to this assumption, as one might imagine scenarios where a neuron is sensitive to different inputs for different tasks or contexts so that a grand average might not be correct? Or, the functional number of inputs driving the output might be relatively low or otherwise highly fluctuating and less easily averaged over.

(3) On the cerebellar example, it is nice that the Bayesian example provides a narrower PF-CF interval for plasticity than the classical rules, but the window is not nearly as narrow as the Suvrathan et al. 2016 paper cited by the authors. Maybe this is something special about that system having well-defined, delayed feedback, but (optional) further comments or insights would be welcome if available.

(4) In the discussion, I appreciated the comparison with the Deneve work which has fast and slow feedback components. I was curious whether, although non-local, there were also conceptual similarities with FORCE learning in which there is also an immediate correction of activity through fast changing of synaptic weights, which then aids the slow long-term learning of synaptic weights.

Reviewer #2 (Public review):

Summary:

Bricknell and Latham investigate the computational benefits of a dual-learning algorithm that combines a rapid, millisecond-scale weight adjustment mechanism with a conventional, slower gradient descent approach. A feedback error signal drives both mechanisms at the synaptic level.

Strengths:

Integrating these two learning timescales is intriguing and demonstrates improved performance compared to classical strategies, particularly in terms of robustness and generalization when learning new target signals.

Weaknesses:

The biological plausibility and justification for the proposed rapid learning mechanism require further elaboration and supporting mechanistic examples.