Bayesian inference is facilitated by modular neural networks with different time scales.

Area of Science:

• Computational neuroscience
• Neural networks
• Bayesian inference

Background:

• Animals, including humans, use Bayesian inference to process noisy, time-varying environmental data.
• The brain's mechanism for acquiring and representing prior distributions for Bayesian inference via neural activity remains unclear.

Purpose of the Study:

• To elucidate how neural activities represent prior distributions for Bayesian inference.
• To investigate the role of network structure and neuronal time scales in Bayesian inference accuracy.

Main Methods:

• Simulated neural networks with modular structures (fast and slow modules) and uniform time scales.
• Training networks to learn and represent prior distributions from noisy inputs.
• Analyzing the emergent network properties and their functional roles.

Main Results:

• Modular networks with fast and slow modules outperformed uniform time scale networks in Bayesian inference.
• Slow sub-modules within the modular network effectively represented prior information, including means and variances.
• The slow-fast modular structure and role specialization emerged spontaneously during training.

Conclusions:

• Modular neural networks with hierarchical time scales are adept at representing prior distributions for Bayesian inference.
• Slow neural modules are critical for integrating information and representing prior statistics.
• This finding provides insight into brain information processing and the significance of slow neural dynamics.