— Omar Montasser 2019/06/10 15:06
This is an experiment where volunteers spend (at most) 5 minutes describing a research problem and plan of attack.
The goal is to exhibit a tiny seed (carrot?) from which a research collaboration can grow.
Omar Montasser - Problems on Robust Learnability:
For one-hidden layer neural networks with threshold/ReLU activations, are there gaps between proper/improper robust learning algorithms in terms of sample complexity?
Given some robustness criterion (i.e. a threat model), can we efficiently robustly PAC learn linear classifiers in the realizable setting?
Empirical/Theoretical analysis of boosting/bagging/smoothing procedures in the robust learning setting, going beyond robust empirical risk minimization.