Dynamical Isometry is Achieved in Residual Networks in a Universal Way for any Activation Function
September 24, 2018 Β· Declared Dead Β· π International Conference on Artificial Intelligence and Statistics
"No code URL or promise found in abstract"
Evidence collected by the PWNC Scanner
Authors
Wojciech Tarnowski, Piotr WarchoΕ, StanisΕaw JastrzΔbski, Jacek Tabor, Maciej A. Nowak
arXiv ID
1809.08848
Category
stat.ML: Machine Learning (Stat)
Cross-listed
cs.LG
Citations
40
Venue
International Conference on Artificial Intelligence and Statistics
Last Checked
3 months ago
Abstract
We demonstrate that in residual neural networks (ResNets) dynamical isometry is achievable irrespectively of the activation function used. We do that by deriving, with the help of Free Probability and Random Matrix Theories, a universal formula for the spectral density of the input-output Jacobian at initialization, in the large network width and depth limit. The resulting singular value spectrum depends on a single parameter, which we calculate for a variety of popular activation functions, by analyzing the signal propagation in the artificial neural network. We corroborate our results with numerical simulations of both random matrices and ResNets applied to the CIFAR-10 classification problem. Moreover, we study the consequence of this universal behavior for the initial and late phases of the learning processes. We conclude by drawing attention to the simple fact, that initialization acts as a confounding factor between the choice of activation function and the rate of learning. We propose that in ResNets this can be resolved based on our results, by ensuring the same level of dynamical isometry at initialization.
Community Contributions
Found the code? Know the venue? Think something is wrong? Let us know!
π Similar Papers
In the same crypt β Machine Learning (Stat)
R.I.P.
π»
Ghosted
R.I.P.
π»
Ghosted
Distilling the Knowledge in a Neural Network
R.I.P.
π»
Ghosted
Layer Normalization
R.I.P.
π»
Ghosted
Dropout as a Bayesian Approximation: Representing Model Uncertainty in Deep Learning
R.I.P.
π»
Ghosted
Domain-Adversarial Training of Neural Networks
R.I.P.
π»
Ghosted
Deep Learning with Differential Privacy
Died the same way β π» Ghosted
R.I.P.
π»
Ghosted
Language Models are Few-Shot Learners
R.I.P.
π»
Ghosted
PyTorch: An Imperative Style, High-Performance Deep Learning Library
R.I.P.
π»
Ghosted
XGBoost: A Scalable Tree Boosting System
R.I.P.
π»
Ghosted