Equi-normalization of Neural Networks
This addresses the problem of over-parametrization for neural network practitioners by offering an alternative to normalization methods, though it appears incremental compared to existing techniques.
The paper tackles the over-parametrization in neural networks by introducing an iterative method to minimize the L2 norm of weights, which improves test accuracy when interleaved with SGD, achieving competitive results on CIFAR-10 and ImageNet with ResNet-18.
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.