Equi-normalization of Neural Networks

Pierre Stock (1, 2) , Benjamin Graham (1) , Rémi Gribonval (2) , Hervé Jégou (1)

1 FAIR - Facebook AI Research [Paris]
2 PANAMA - Parcimonie et Nouveaux Algorithmes pour le Signal et la Modélisation Audio

Modern neural networks are over-parametrized. In particular, each rectified linear hidden unit can be modified by a multiplicative factor by adjusting input and output weights, without changing the rest of the network. Inspired by the Sinkhorn-Knopp algorithm, we introduce a fast iterative method for minimizing the L2 norm of the weights, equivalently the weight decay regularizer. It provably converges to a unique solution. Interleaving our algorithm with SGD during training improves the test accuracy. For small batches, our approach offers an alternative to batch-and group-normalization on CIFAR-10 and ImageNet with a ResNet-18.

Domaines: Intelligence artificielle [cs.AI]

Pierre Stock, Benjamin Graham, Rémi Gribonval, Hervé Jégou. Equi-normalization of Neural Networks. ICLR 2019 - Seventh International Conference on Learning Representations, May 2019, New Orleans, United States. pp.1-20. ⟨hal-02050408⟩