Early local objective training methods such as layer-wise pretraining typically underperformed compared to end-to-end training.