Synaptic plasticity as Bayesian inference.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• Rapid learning is crucial for survival, but faces challenges due to noisy sensory data.
• Synaptic plasticity, the ability of synapses to strengthen or weaken over time, is fundamental to learning and memory.
• Current models often simplify the complex process of synaptic weight adjustment.

Purpose of the Study:

• To propose a novel Bayesian inference framework for synaptic plasticity.
• To hypothesize how synapses represent and utilize uncertainty in weight estimation.
• To explain the variability in postsynaptic potential (PSP) size.

Main Methods:

• Formulating two core hypotheses about synaptic uncertainty representation and utilization.
• Casting synaptic plasticity as a Bayesian inference problem.
• Generalizing existing synaptic plasticity rules.

Main Results:

• Synapses are hypothesized to estimate weights with associated uncertainty (error bars).
• Higher uncertainty in weight estimates leads to higher learning rates.
• Synaptic uncertainty is communicated via variability in postsynaptic potential size.

Conclusions:

• The proposed Bayesian framework provides a normative view of learning and synaptic plasticity.
• This model offers a unified explanation for synaptic learning rate modulation and PSP variability.
• The hypotheses generate falsifiable predictions for experimental validation.