Statistical inference on representational geometries
Heiko Schütt, Alexander D. Kipnis, Nikolaus Kriegeskorte, Columbia University, United States; Jörn Diedrichsen, Western University, Canada
Posters 3 Poster
Pacific Ballroom H-O
Sat, 27 Aug, 19:30 - 21:30 Pacific Time (UTC -7)
Representational similarity analysis is a versatile method for comparing high dimensional models and neural recording data to each other. Here, we introduce a comprehensive new set of methods for statistical model comparison based on predictions of representational geometries. The inference can handle flexible parametrized models and can treat both subjects and conditions as random effects, such that conclusions generalize to the respective populations of subjects and conditions. With crossvalidated representational distance estimators and metric whitened model evaluators, the power for model comparisons approximates that of likelihood-based inference, but rank-based model evaluation is also supported. We validate the inference methods using extensive simulations with deep neural networks and resampling of calcium imaging and functional MRI data. Results demonstrate that the methods are valid and conclusions generalize correctly. These data analysis methods are available in an open-source Python toolbox.