In the context of deep learning, how does the degradation problem affect training efficiency and model complexity in neural networks?

When a neural network gets deeper, its accuracy can actually drop instead of improve, a phenomenon called the degradation problem. This happens because extra layers make the gradient signals weaker, so the optimizer has trouble learning the right weights and training slows down. As a result, you need more epochs or more complex architectures like residual connections to keep training efficient. For example, a plain 50‑layer CNN may perform worse than a 20‑layer one unless shortcut connections are added. Thus, the degradation problem forces us to either simplify the model or use tricks that keep training effective.