Toward the Bayesian brain: a generative model of information transmission by vestibular sensory neurons.

The relative accessibility and simplicity of vestibular sensing and vestibular-driven control of head and eye movements has made the vestibular system an attractive subject to experimenters and theoreticians interested in developing realistic quantitative models of how brains gather and interpret sense data and use it to guide behavior. Head stabilization and eye counter-rotation driven by vestibular sensory input in response to rotational perturbations represent natural, ecologically important behaviors that can be reproduced in the laboratory and analyzed using relatively simple mathematical models. Models drawn from dynamical systems and control theory have previously been used to analyze the behavior of vestibular sensory neurons. In the Bayesian framework, which is becoming widely used in cognitive science, vestibular sense data must be modeled as random samples drawn from probability distributions whose parameters are kinematic state variables of the head. We show that Exwald distributions are accurate models of spontaneous interspike interval distributions in spike trains recoded from chinchilla semicircular canal afferent neurons. Each interval in an Exwald distribution is the sum of an interval drawn from an Exponential distribution and a Wald or Inverse Gaussian distribution. We show that this abstract model can be realized using simple physical mechanisms and re-parameterized in terms of the relevant kinematic state variables of the head. This model predicts and explains statistical and dynamical properties of semicircular canal afferent neurons in a novel way. It provides an empirical foundation for realistic Bayesian models of neural computation in the brain that underlie the perception of head motion and the control of head and eye movements.