The Limits of Post-Selection Generalization
summary
Examines limits of post-selection generalization in data analysis, focusing on connections between generalization and parameters like sample size and adaptivity. Presents formal results about when post-hoc selection of analyses prevents valid generalization, using theoretical tools from learning theory and privacy-related frameworks (e.g., differential privacy concepts) to characterize these limits.
AI-written summary · about
related talks
53mGeneralization insights from actual cognition
42mWhat (Doesn't) Make Learning Algorithms Generalize?
10mModern Paradigms in Generalization
45mStudying Generalization in Deep Learning via PAC-Bayes
01h 07mDoes generalization imply accuracy on the line? A new look at robust generalization
01h 10mModern paradigms of generalization, the heliocentric model of Aristarchus,...
38mRepresentation via Representations: Improving Generalization to Unseen Populations
35mProbably Approximately Precision and Recall Learning