We're launching a Kaggle competition on the Abstraction & Reasoning Corpus (ARC)! 3 months duration, $20,000 in prizes. Good luck :)
kaggle.com/c/abstraction-…
kaggle.com/c/abstraction-…
Huge thanks to the wonderful folks at Kaggle who worked on this! :)
ARC is a set of IQ-like tasks that:
1) don't reuse knowledge other than "Core Knowledge" (basic set of knowledge primitives that even v. young children possess)
2) Are readily solvable by humans, without training or instructions
3) Are out of reach for current AI... or are they?
1) don't reuse knowledge other than "Core Knowledge" (basic set of knowledge primitives that even v. young children possess)
2) Are readily solvable by humans, without training or instructions
3) Are out of reach for current AI... or are they?
I hope this challenge will help the rise of alternative approaches to program synthesis that leverage human cognitive priors & focus on abstraction & generalization -- instead of the usual "learn the manifold of the data using a dense sampling of it, then do local interpolation".
