NeurIPS 2019
Sun Dec 8th through Sat the 14th, 2019 at Vancouver Convention Center
Paper ID:3809
Title:A Mean Field Theory of Quantized Deep Networks: The Quantization-Depth Trade-Off

The paper provides a mean-field analysis of infinitely wide neural networks with quantized activations, proposing a relation between the choice of initialization hyper-parameters and the maximal depth by primarily by considering how correlations between two inputs propagate through the network at initialization as well as numerical stability issues. All reviewers agree that it is a good contribution.