Optimal Decoding of Dynamic Stimuli by Heterogeneous Populations of Spiking Neurons: A Closed-Form Approximation.

Area of Science:

• Computational Neuroscience
• Systems Neuroscience
• Information Theory

Background:

• Neural decoding, often framed as dynamic state estimation from point-process observations, is computationally challenging.
• Existing numerical sampling methods offer practical solutions for neural data decoding but lack theoretical depth for understanding optimal encoding/decoding strategies.

Purpose of the Study:

• To develop an analytically tractable Bayesian approximation for optimal filtering in neural systems.
• To facilitate the analysis of optimal neural encoding strategies, especially in scenarios deviating from uniform coding assumptions.
• To provide a theoretical framework for understanding sensory neural populations characterized by parameter distributions.

Main Methods:

• Formulated neural decoding as dynamic state estimation using point-process observations.
• Developed an analytically tractable Bayesian approximation to optimal filtering.
• Utilized continuous distributions to approximate large neural populations, simplifying filter complexity.
• Compared the approximation's accuracy with particle filtering.

Main Results:

• The proposed Bayesian approximation offers significant analytical tractability for neural decoding.
• The filter's complexity is independent of population size, enabling optimization of population parameters.
• Numerical comparisons validate the approximation's quality against particle filtering.
• The analytic framework yields insights not easily obtainable from numerical methods.

Conclusions:

• The developed analytic framework provides valuable theoretical insights into optimal neural encoding and decoding.
• This approach is consistent with biological observations regarding sensory cell tuning.
• The method offers a computationally efficient and theoretically robust alternative to numerical sampling for neural decoding.