Multi-voxel similarity analysis for rapid learning designs: challenges and possible solutions?

Date

May 11, 2017

Time

4:00 PM - 5:00 PM

Speaker

Dr. Hannes Ruge

Affiliation

TUD, Allgemeine Psychologie

Series

TUD NIC Kolloquium

Language

en

Main Topic

Psychologie

Other Topics

Psychologie

Description

Multi-voxel similarity analysis based on simple correlation measures is known to be biased due to even moderate collinearity among model regressors. By strict randomization of experimental conditions across trials and subjects, it is possible to get rid of condition-specific biases. However, in learning experiments this cannot be done due to the inherent sequential dependency of experimental conditions (i.e. late vs. early in learning). A possible way out could be to use classifier approaches instead of correlation-based pattern similarity. Yet, classifiers rely on a sufficient number of training and testing exemplars. Unfortunately, exactly this is not given in rapid learning experiments with only few trials per learning item. I am presenting attempts to nevertheless apply multi-voxel pattern similarity analysis and discuss possible ways how to assess and tame the bias introduced by regressor collinearity.

Links

• Livestream

Last modified: May 9, 2017, 12:48:08 PM