Learning probability distributions of sensory inputs with Monte Carlo predictive coding.

Area of Science:

• Computational neuroscience
• Cognitive science
• Machine learning

Background:

• The brain is hypothesized to use probabilistic generative models for sensory interpretation.
• Distinct frameworks like predictive coding and neural sampling explain separate aspects of this process.
• Variational filtering previously integrated these frameworks, introducing neural sampling to predictive coding.

Purpose of the Study:

• To introduce Monte Carlo predictive coding (MCPC), a variant of variational filtering for static inputs.
• To demonstrate how MCPC integrates predictive coding and neural sampling for learning generative models.
• To show MCPC's ability to infer posterior distributions and generate sensory inputs.

Main Methods:

• Developed a novel neural network model: Monte Carlo predictive coding (MCPC).
• Utilized variational filtering principles adapted for static inputs.
• Integrated predictive coding with neural sampling mechanisms.

Main Results:

• MCPC learns precise generative models through local computation and plasticity.
• Neural dynamics in MCPC infer posterior distributions of latent states.
• MCPC can generate likely sensory inputs in the absence of actual inputs.
• The model captures experimental observations of neural activity variability during perceptual tasks.

Conclusions:

• MCPC successfully combines predictive coding and neural sampling into a unified framework.
• The model accounts for neural data previously explained by individual frameworks.
• MCPC offers a potential explanation for optimal sensory interpretation and neural variability in the brain.