These limitations of XAI are the key point really, and I could have stopped here, but I gave some more details. I talked...

These limitations of XAI are the key point really, and I could have stopped here, but I gave some more details. I talked about stability (gene signatures remain predictive but change when you search for them on different subsets of the same data), and mentioned that both stability selection procedures and structured sparsity approaches improve stability. (Cue in a long list of references to my own work. I spared the audience, and I'll spare you.) And then I mentioned existing methods for (1) doing inference, meaning being able to provide e.g. p-values or confidence intervals on machine learning models: I cited post-selection inference (PSI) and inference for conditional feature importance (CFI)(2) controlling the false discovery rate (see thread above on knockoffs). These methods are, in my opinion, sadly understudied--Angelo Reyero-Lobo, Pierre Neuvial, and Bertrand Thirion only showed how to perform inference with CFI earlier this year.6/6#machineLearning #genomics