Project Details
Projekt
Visual explanations for statistical tests and statistical tests for visual explanations with application to imaging genetics
Subject Area
Medical Informatics and Medical Bioinformatics
Image and Language Processing, Computer Graphics and Visualisation, Human Computer Interaction, Ubiquitous and Wearable Computing
Statistics and Econometrics
Image and Language Processing, Computer Graphics and Visualisation, Human Computer Interaction, Ubiquitous and Wearable Computing
Statistics and Econometrics
Term
since 2022
Project identifier
Deutsche Forschungsgemeinschaft (DFG) - Project number 459422098
Recently, progress has been made in utilizing the ability of deep neural networks to extract relevant information from data in statistical tests to test whether two sets of potentially non-tabular examples are different in distribution (Kirchler et al., 2020). While tests based on deep learning achieve high power to detect the presence of population-level differences between the two sets of examples (e.g. images), they do not answer the question, which features (e.g. pixels) make the populations different. This black-box nature hinders biomedical applications that aim to identify features associated with populations like carriers or non-carriers of a genetic mutation. Explainable AI identify important features by highlighting features in one sample that explain the prediction of a trained neural network, but provide limited information about the population.One goal of this project is to develop explainable AI techniques for statistical tests, including the aforementioned deep learning two-sample test. To do so, we will adapt Layer-wise Relevance Propagation (Bach et al., 2015) to provide visual explanations of population-level differences in the deep learning representation space. This will allow us to understand and verify statistical testing results and e.g. visualize differences between two populations. We will extend these explanations to deep-learning-based conditional independence tests that test for (in)dependence between a structured variable (e.g. image) and a scalar variable (e.g. genetic score) when conditioning on a set of covariates and confounders. We will apply our explanation methods to the genetic analysis of image phenotypes (e.g. MRI or eye fundus images) to visualize putatively heritable patterns. A second goal of this project is to bring the formalism of statistical analysis to the explainable AI domain. To this end, we will develop novel statistical tests for the analysis of a set of individual explanations computed with Layer-wise Relevance Propagation. First steps towards a population-wide analysis of individual explanations have been done by Samek and colleagues in the recent work on Spectral Relevance Analysis (SpRAy) (Lapuschkin et al., 2019). While SpRAy computes prototypical explanations from a set of individual explanation heatmaps by clustering, we will leverage statistical methods to analyze the set of heatmaps, e.g., test whether the model applies different classification strategies to different groups (e.g., men and women, carriers/non-carriers of a genetic mutation) or whether a specific explanation pattern covaries with the genotype. These statistical tools will not only foster trust in the explanations obtained in biomedical applications, but will bring forward explainable AI as a whole.
DFG Programme
Research Units
