0f4d1fc085b7504c140e66bb26ed8842-Paper-Conference.pdf

Neural Information Processing Systems 

Empirical works show that for ReLU neural networks (NNs) with small initialization, input weights of hidden neurons (the input weight of a hidden neuron consists oftheweight fromitsinputlayertothehidden neuron anditsbiasterm) condense ontoisolated orientations.

Similar Docs  Excel Report  more

TitleSimilaritySource
None found