Bayesian computation emerges in generic cortical microcircuits through spike-timing-dependent plasticity.

Area of Science:

• Computational Neuroscience
• Systems Neuroscience
• Theoretical Neuroscience

Background:

• The computational principles of neural networks and the role of spike-timing dependent plasticity (STDP) remain largely unknown.
• Soft winner-take-all (WTA) circuits, characterized by mutual inhibition among pyramidal neurons via interneurons, are prevalent in cortical microcircuits.

Purpose of the Study:

• To elucidate how STDP and neuronal excitability contribute to Bayesian computation in neural networks.
• To investigate the emergence of implicit generative models and priors within synaptic weights and neuronal adaptation.

Main Methods:

• Theoretical analysis of neural network dynamics.
• Computer simulations of neuronal activity and synaptic plasticity.
• Investigation of STDP's capacity to approximate Expectation Maximization.

Main Results:

• STDP combined with activity-dependent excitability induces Bayesian computation in WTA circuits.
• Neuronal adaptation generates priors, while STDP shapes synaptic weights to form implicit generative models.
• STDP approximates Expectation Maximization for fitting generative models to high-dimensional spike inputs.

Conclusions:

• Spontaneous neural activity and trial-to-trial variability are essential for representing probability distributions, not static codes.
• Networks of Bayesian computation modules offer a novel framework for distributed information processing in the cortex.