Dynamical Isometry is Achieved in Residual Networks in a Universal Way for any Activation Function.
Wojciech TarnowskiPiotr WarcholStanislaw JastrzebskiJacek TaborMaciej A. NowakPublished in: CoRR (2018)
Keyphrases
- activation function
- network size
- neural network
- feed forward neural networks
- feed forward
- artificial neural networks
- hidden layer
- back propagation
- radial basis function
- chaotic neural network
- network structure
- learning rate
- training phase
- vector field
- neural nets
- feedforward neural networks
- bayesian networks
- multilayer perceptron
- hidden nodes
- rbf neural network
- training data
- decision making
- social networks