Uniform convergence may be unable to explain generalization in deep learning

Vaishnavh Nagarajan, J. Zico Kolter


This paper proposes an argument against the use of uniform convergence based generalization bounds to explain why overparameterized deep networks generalize well. , they do so by failing the tightest (algorithm, distribution)-dependent uniform convergence bound.

Background Information about the problem

Main Experiment

Main Contributions

Our Two Cents

References and Further reads