Estimating entropy rates with Bayesian confidence intervals

Area of Science:

• Computational neuroscience
• Information theory
• Dynamical systems

Background:

• Entropy rate quantifies uncertainty in dynamical systems.
• Estimating entropy rate in neural data is challenging due to model biases.
• Neural information encoding is linked to this uncertainty.

Purpose of the Study:

• Develop a direct and accurate estimator for the entropy rate from experimental data.
• Improve upon existing methods for bias and convergence speed.
• Provide a robust framework for analyzing neural information processing.

Main Methods:

• Applied model weighting principle from lossless data compression (Minimum Description Length).
• Developed a direct estimator for entropy rate.
• Utilized Monte Carlo techniques for Bayesian confidence intervals.

Main Results:

• The new estimator shows significantly less bias compared to existing methods.
• The estimator converges faster in simulations.
• Successfully applied to estimate information rates in neural responses to stimuli.

Conclusions:

• The Minimum Description Length-based approach provides a superior method for entropy rate estimation in neural systems.
• This technique enhances the analysis of information encoding in neuroscience.
• Offers a more reliable tool for interpreting complex neural dynamics.