Natural statistics of head roll: implications for Bayesian inference in spatial orientation

Area of Science:

• Neuroscience
• Perception
• Computational modeling

Background:

• Bayesian models are used to understand multisensory integration in spatial orientation.
• A previously proposed Gaussian prior model successfully explained perceptual phenomena like subjective visual vertical and body tilt.
• Natural head motion statistics are increasingly recognized as non-Gaussian, prompting an investigation into their impact on Bayesian models.

Purpose of the Study:

• To investigate the performance of a Bayesian model for spatial orientation using a non-Gaussian prior.
• To compare the explanatory power of Gaussian versus non-Gaussian priors for perceptual data.
• To explore the computational and neurophysiological basis for the choice of prior distribution in vertical perception.

Main Methods:

• Experimentally characterized natural head orientation statistics, quantifying them as a t-location-scale distribution.
• Compared the performance of the original Bayesian model with a Gaussian prior against variants using a t-distributed prior.
• Evaluated model performance on previously published data for subjective visual vertical and subjective body tilt tasks.

Main Results:

• Natural head orientation statistics exhibit long tails, consistent with a t-location-scale distribution.
• Bayesian model variants employing a t-distributed prior performed substantially worse than the original model with a Gaussian prior.
• The Gaussian prior demonstrated a superior account of perceptual observations compared to non-Gaussian alternatives.

Conclusions:

• A Gaussian prior centered on upright head orientation remains the most effective for explaining spatial orientation perception.
• Despite natural head motion statistics being non-Gaussian, a Gaussian prior offers a better fit for perceptual data.
• The precision-accuracy trade-off associated with the Gaussian prior likely underlies its effectiveness across the tilt range.