I didn't really get why multi-scale aggregation would solve adversarial examples until it just occurred to me that one exploitable difference between a real reward and Goodhart points is that real…

Posters: jd_pressman. jd_pressman

4 likes recorded by the source corpus.