Population codes of prior knowledge learned through environmental regularities.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• The brain infers environmental states from noisy data using prior knowledge.
• Efficient Bayesian inference is observed in humans, but neural mechanisms for acquiring and representing prior knowledge are unclear.
• Experimental investigation of prior knowledge acquisition is challenging due to long time scales.

Purpose of the Study:

• To investigate how neural populations learn and represent prior probabilities.
• To provide empirical predictions for future neuroscience experiments.
• To model neural responses to expected and unexpected stimuli.

Main Methods:

• Utilized a neural network model trained on orientation classification and estimation tasks.
• Manipulated prior probabilities of stimuli during model training.
• Analyzed neural population activity and tuning properties.

Main Results:

• Neural network populations learned to represent and incorporate prior knowledge with trial-to-trial accuracy feedback only.
• Identified a novel mechanism where neuronal activation thresholds adapt based on stimulus prior probability.
• Observed denser and narrower tuning curves for more likely stimuli, enhancing information processing.

Conclusions:

• Neural networks can learn and represent prior knowledge without explicit probabilistic feedback.
• Adaptive activation thresholds and optimized tuning curves are key mechanisms for utilizing prior knowledge.
• Findings explain experimental observations and offer testable predictions for neural representations of prior knowledge.