Ensemble normalization for stable training

1) The document proposes an ensemble normalization method that combines two existing normalization techniques to improve stable training of neural networks. 2) Existing normalization methods like batch, instance, and group normalization have limitations when used individually. The proposed ensemble method calculates the addition and division of two normalization methods. 3) Experimental results on semantic segmentation and image generation tasks show the ensemble method improves performance over single normalization baselines. It allows different features to be learned compared to existing techniques.