Sampling-based Bayesian inference in recurrent circuits of stochastic spiking neurons.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Theoretical Neuroscience

Background:

• Cortical circuits exhibit significant neuronal response variability, often near Poisson statistics.
• Abundant recurrent connections are a hallmark of cortical circuitry.
• The interplay between spiking variability and recurrent connectivity in sensory processing remains unclear.

Purpose of the Study:

• To develop a theoretical framework explaining how cortical spiking variability and recurrent connections enable optimal Bayesian inference.
• To elucidate the role of these neural properties in sensory representation and information processing.
• To provide an experimentally testable prediction for the proposed framework.

Main Methods:

• Construction of a theoretical framework integrating neuronal spiking variability and recurrent network dynamics.
• Modeling of Bayesian inference through flexible sampling from posterior stimulus distributions.
• Illustration of the framework's application to hierarchical and parallel stimulus representations.

Main Results:

• The framework demonstrates that cortical features facilitate optimal sampling-based Bayesian inference.
• Recurrent connections are shown to encode internal models of the external world.
• Poissonian spike variability enables flexible sampling from complex stimulus distributions.

Conclusions:

• The integration of neuronal variability and recurrent connections supports efficient sensory information processing in the cortex.
• The proposed model offers a mechanistic explanation for how the brain performs Bayesian inference.
• Internally generated differential correlations serve as a neural signature for network sampling, linked to stored priors.