Exploring deep neural networks via layer-peeled model: Minority collapse in imbalanced training