Slide from layer 1 to 5 and watch the values drift and spread out, so each layer's input keeps shifting. Batch normalization is built to fix this.
Each layer drifts and stretches what it passes on. Deeper layers are chasing a target that keeps moving underneath them.
