A Bayesian model of auditory performance

The auditory system is large and complex. Most auditory models are decidedly deterministic, except for a spiking model. We'll never have enough tests to fully characterize it. Can Bayesian modeling help, summarizing what we know, and providing probability distributions for the parts we do not know?

Malcolm Slaney will provide an introduction to Bayesian modeling and describe a first-ever probabilistic model of speech perception performance in noise. There are a number of disparate tests to characterize our auditory system. Earlier this quarter we’ve talked about DNN and other optimization techniques to fit biophysical models of the auditory system to test data. But what do we do when we don’t have all the necessary tests? Can we reason about the missing data, generating maximum a posteriori estimates, and estimating our degree of uncertainty? The uncertainty might be important because it gives us confidence in our models, and perhaps also tells us where to test next.

This is work in progress, so lots of questions are appropriate. After an introduction to Bayesian modeling, I’ll present initial clinical results, and we can discuss what the models tell us, and where to go next. This excellent overview will help get you up to speed: Bayesian statistics and modeling.



Who: Malcolm Slaney
What: A Bayesian Model of Auditory Performance
When: Friday March 15th at 10:30AM
Where: CCRMA Stage <<< New location this week
Why: Can we can measure uncertainty?