Calibrating Bayesian Decoders of Neural Spiking Activity

Area of Science:

• Systems Neuroscience
• Computational Neuroscience
• Neuroscience

Background:

• Accurate decoding of external variables from neural activity is crucial in systems neuroscience.
• Bayesian decoders are widely used for probabilistic estimation of neural data.
• Overconfidence in traditional Bayesian decoder predictions is a common issue.

Purpose of the Study:

• To identify and characterize overconfidence in traditional Bayesian decoders.
• To investigate Bayesian decoding with latent variables for improved calibration.
• To propose methods for correcting miscalibration in neural decoding.

Main Methods:

• Analysis of neural recordings from multiple brain regions and species (monkeys, mice, rats).
• Application of Bayesian decoding techniques, including those with latent variables.
• Characterization of prediction overconfidence and development of post hoc correction methods.

Main Results:

• Demonstrated overconfidence in traditional Bayesian decoders across diverse neural decoding tasks.
• Showed that Bayesian decoding with latent variables can improve calibration.
• Identified the need for additional post hoc correction for optimal calibration.

Conclusions:

• Properly calibrated Bayesian decoders are essential for accurate probabilistic population coding.
• Improved calibration can lead to more reliable brain-machine interfaces.
• Addressing decoder overconfidence enhances the trustworthiness of decoded neural information.