How to initialize weights in the presence of skip connections?

Question

Weight initialization is an important parameter for success in large networks, in the absence of techniques such as batch normalization that reduces their impact.

There are known initialization techniques such as "Gloriot" or "He", but they are based on a "standard" feed-forward architecture model, where outputs of a layer are the inputs of the next one.

Is there any analysis (theoretical or experimental) of weight initialization in the presence of skip connections?

Answer 1

There's a paper where they could train a ResNet without batchnorm, using only a carefully chosen way of initializing the network: https://arxiv.org/pdf/1901.09321.pdf