Search papers, labs, and topics across Lattice.
Department of Computer and Information Sciences, University of Pennsylvania
2
0
3
Zero model disagreement is achievable with simple techniques like stacking, boosting, and architecture search, converging as you increase the number of models, iterations, or architecture size.
Even with the option to abstain, distribution-agnostic learners in adversarial settings are provably worse than those with distribution knowledge, but a new "certificate dimension" approach offers a path to tighter bounds.
