Keynotes | K&Ts | GACs | Talks | Posters | Search

Poster B in Poster Session B: Tuesday, August 4, 2:00 – 3:45 pm, Kimmel Center, Shorin & Rosenthal Rooms

Estimating the generative model underlying stochastic processes drives auditory perception

Clémentine Lévy-Fidel¹, Jasmin Stein², Alejandro Tabas¹; ¹Basque Center on Cognition, Brain and Language, ²Max Planck Institute for Human Cognitive and Brain Sciences

Presenter: Clémentine Lévy-Fidel

Making sense of noisy sensory information is challenging. The brain makes use of what it has learnt in the past to aid this process. The use of prior information in perception has been traditionally formalised within the Bayesian framework, which proposes that the brain estimates the generative model underlying sensory inputs to compute optimal prior beliefs on their latent causes. Correct priors yield posterior beliefs with lower uncertainty, potentially leading to faster and sharper perceptual decisions. Previous work investigated the effect of priors in perceptual decision-making using tasks where the statistics of the generative model were either explicitly communicated to participants or held fixed across trials. In such settings, participants do not need to estimate the parameters of the generative model: they can rely on a fixed prior, bypassing the inference process that characterises perception in dynamically evolving environments. Here we investigate whether participants can estimate and deploy internal models of an ambiguous environment during perception. We designed a stochastic auditory paradigm that requires estimating the parameters of the generative model underlying the data. The paradigm is an extension of the classical oddball, where regularly presented standards are sampled from a linear Gaussian dynamic process. Participants are instructed to detect the deviant tone, which is sampled from a process with a slightly different mean. We recorded reaction times in participants responses and enquired whether they estimated the parameters of the generative model underlying the sampling of the standards. We used a Kalman filter to model optimal estimation of the generative model. Reaction times correlated closely with the optimal likelihood estimates derived from the Kalman filter, suggesting that participants learn the statistical structure of the sensory environment, exploit this knowledge to generate predictions, and respond with a speed that reveals surprise about the incoming stimulus.

Topic Area: Auditory, Speech & Language Processing