Automatically discovering and validating interpretable features in a model without requiring human-labeled concepts.