Auditory reverse correlation is an experimental paradigm which allows researchers to reveal the acoustic features that are used by listeners in an auditory task. It can be thought of as a sort of "ear-tracking" technique, based on purely behavioral data. It relies on a stimulus-response model, fitted using advanced machine learning techniques (penalized regression), to produce an instant picture of a participant’s listening strategy in a given context.
The auditory revcorr method (developed as a MATLAB toolbox by our group: https://github.com/aosses-tue/fastACI) is already fully operational and has yielded conclusive results in previous studies. It is, however, very time-consuming (≈ 3 h/participant). The objective of this project will be to improve the efficiency of the estimation process, in order to reduce the duration of the experiment.
On-line Bayesian optimization is an appealing solution for reducing the number of trials in a revcorr experiment. It would consist in replacing the current “linear” offline paradigm (all stimuli generated prior to the testing) with an adaptive approach, in which the next stimulus to be presented depends on the data already collected. This approach has been successfully used for psychometric function estimation and EEG protocols. In a revcorr experiment, online stimulus generation should be a tremendous time saver compared to the current approach where stimuli are independent from the data.
The purpose of this project is to develop an efficient module for the Bayesian optimization process (in MATLAB) then to test the efficiency of the approach on simulated and real listeners.