Measuring Per-Unit Interpretability at Scale Without Humans

Open in new window