Date: Wednesday, 16.10.2024 15:20-17:00 CET
Location: Building S1|15 Room 133
Abstract:
Interesting stimulus spaces are often high-dimensional, which makes it infeasible to exhaustively measure perceptual decisions. Exploring such a space therefore requires adaptive experimental methods. Current adaptive methods are either not suited to classical, well-characterized psychophysical tasks, or do not scale well to more than four dimensions. We propose a particle filtering approach to approximate the posterior distribution of a multidimensional psychometric function with lapses. Even for high dimensions this method is fast enough to update the posterior between trials (order of 1 sec). The posterior distribution allows one to select the next stimulus in the experiment to maximise expected information gain. In simulations, we show that entropy decreases two to three times faster compared to sampling stimuli randomly for a 15-dimensional feature space. We test the algorithm in simulations up to 50 dimensions and find that it is still fast and reliable. We validate the algorithm in a human experiment on facial gender categorization. For faces of the Chicago Face Database, we compute Active Appearance Model features (around 500 000 dimensions), perform dimensionality reduction to 15 dimensions using PCA, and then create a pool of new face images by morphing between faces in the 15-dimensional space. Human participants label the faces as “male” or “female”. The adaptive method is more than twice as efficient as random sampling in terms of entropy minimization, but predictive performance shows mixed results. These results are consistent with the notion that participants change their decision function in an adaptive experiment compared to non-adaptive stimulus selection. While this behaviour suggests caution in interpreting perceptual decision spaces independently of the task, our method nevertheless allows the measurement of perceptual decision functions in stimulus spaces that were previously infeasible.
