Tag
This paper shows that sequence-pooled normalization in convolutional networks provides global context beyond the receptive field, supplying most of the context needed for sequence labeling and affecting attribution in network ablation studies.