"Debugging model mistakes" is a beautiful phrase, especially if it isn't followed by "is difficult".
Also, the formulation of conceptual bank is definitely something I will keep in mind
---
RT @james_y_zou
Want to know why #AI screws up specific predictions?

Our new #ICML2022 paper uses natural language to explain why a model makes each mistake. Key idea: counterfactual analysis of high-level concepts.

Understanding mistakes is criti…
https://twitter.com/james_y_zou/status/1541452062344417280