Network Plasticity as Bayesian Inference.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Statistical Learning Theory

Background:

• Statistical learning theory suggests brain plasticity, like computation, involves probabilistic inference.
• A concrete model for probabilistic inference in brain plasticity was previously lacking.

Purpose of the Study:

• To propose a model where stochastic synaptic plasticity and spine motility enable cortical networks to perform probabilistic inference.
• To offer an alternative to models based on maximum likelihood estimation.
• To explain how neural networks integrate prior knowledge with experience and generalize information.

Main Methods:

• The proposed model utilizes inherently stochastic features of synaptic plasticity and spine motility.
• It posits that cortical networks sample from a posterior distribution of network configurations.

Main Results:

• The model explains optimal merging of prior knowledge (weight distributions, connection probabilities) with learned experience.
• It accounts for the generalization of learned information to novel situations.
• It demonstrates how neural networks can continuously compensate for unforeseen network disturbances.

Conclusions:

• A new theory of network plasticity is presented, viewing it as probabilistic inference.
• This theory functionally explains previously puzzling experimental data on stochastic synaptic plasticity.