1: \begin{abstract}

2: We introduce a new normalization technique that exhibits the fast

3: convergence properties of batch normalization using a transformation

4: of layer weights instead of layer outputs. The proposed technique

5: keeps the contribution of positive and negative weights to the layer

6: output balanced. We validate our method on a set of standard benchmarks

7: including CIFAR-10/100, SVHN and ILSVRC 2012 ImageNet.

8: \end{abstract}

9: