Probabilistic neural transfer function estimation with Bayesian system identification.

Area of Science:

• Computational Neuroscience
• Systems Neuroscience
• Machine Learning

Background:

• Neural population responses are linked to physical stimuli, traditionally characterized by receptive fields.
• Existing neural system identification models require extensive data, posing challenges due to limited experimental recording times.
• Deep neural networks excel at prediction but often lack uncertainty quantification for neural representations and derived statistics like most exciting inputs (MEIs).

Purpose of the Study:

• To develop a Bayesian system identification approach for predicting neural responses to visual stimuli.
• To investigate the benefits of modeling network weight variability for identifying neural response properties.
• To provide uncertainty estimates for neural representations and derived statistics, enhancing model evaluation and feature interpretation.

Main Methods:

• Employed variational inference to estimate the posterior distribution of model weights from training data.
• Developed a Bayesian system identification framework to predict neural responses and quantify uncertainty.
• Utilized an effectively infinite ensemble generated by the variational method to derive most exciting inputs (MEIs).

Main Results:

• The Bayesian approach achieved higher or comparable neural prediction performance with significantly improved data efficiency compared to Monte Carlo dropout and traditional point-estimate models.
• The method generated an ensemble of models, enabling robust estimation of stimulus-response function uncertainty, which correlated negatively with predictive performance.
• In silico experiments demonstrated that the model generated stimuli driving neuronal activity more effectively than traditional models under data-limited conditions.

Conclusions:

• Bayesian system identification with variational inference offers a data-efficient method for neural response prediction and system characterization.
• Explicitly modeling network weight variability provides crucial uncertainty estimates, aiding in the evaluation and interpretation of neural models and their inferred properties.
• The approach facilitates the identification of meaningful neural response properties with credible intervals, advancing the understanding of sensory systems in data-limited scenarios.