The residual learning framework primarily improves the training of deep neural networks by allowing them to learn unreferenced transformations instead of residuals, making it unnecessary to consider input data from earlier layers.

Residual learning lets a deep network focus on learning the difference between the desired output and the input that comes from the preceding layers, instead of trying to learn the whole mapping from scratch. By predicting this “residual” signal, the network can keep the main signal unchanged and only adjust what is needed, which makes the optimization landscape smoother and gradients easier to propagate. This approach reduces the risk of vanishing gradients and allows very deep models to train effectively, because each block only has to learn a small correction rather than a complete transformation. For example, a 100‑layer network can learn a residual that is close to zero for many layers, letting the earlier layers provide a strong baseline and the later layers refine it. Thus residual learning improves training by simplifying each block’s task and preserving the original information from earlier layers.