mlcommons / training_policies

Issues related to MLPerf™ training policies, including rules and suggested changes
https://mlcommons.org/en/groups/training
Apache License 2.0
93 stars 66 forks source link

Moving away from white-list/black-list? #271

Open bitfort opened 4 years ago

bitfort commented 4 years ago

There is a need to build a "shared intuition" (versus a white/black list) among submitters for what optimizations are allowed. We want some way this alignment can be also be used to reduce confusion/disagreement during post submission review. This could take the form of a "conceptual test", statement of "principles", or something else.

Examples; what parts of minigo can be changed? How can step schedules be compressed?

AI(Victor) - Circulate a document where we can enumerate criteria for and discuss a "conceptual test".