Gary Marcus

“Correlation does not equal causation.” #ArtificialIntelligence #AI #MachineLearninghttps://twitter.com/GaryMarcus/status/1170912860815396864 …

No one in AI/ML would ever say something so silly. Why bring it up? Causation may be established through randomized experiments, or by varying each hypothesized causal factor in turn while holding the rest fixed. 2nd order statistics may then be used to represent causal factors.

Knowing which variables to try controlling for requires already having a causal model in the first place though - otherwise controlling can either uncover/obscure causal factors depending on whether it's a confounder/mediator/blocker

teenagers induce causal models of alternative worlds every time they master a video game. some day AI will be able to do the same, for the real world. cc @yudeapearl @eliasbareinboim

It's one thing to teach AI how to reason about counterfactuals given a causal model (e.g. do-calculus @yudapearl) - seems like another (much harder) problem to generate these causal models from scratch like the human brain can seemingly do super easily from a young age