Which of the following statements about identity mapping in deep models are true? Select all that apply.

Identity mapping is a key idea in residual networks, where a shortcut connection simply adds the input to the output of a few layers; this keeps the signal from being distorted too much and lets gradients flow backwards without shrinking. Because the shortcut passes the raw input forward, the network can learn small deviations from the identity rather than building everything from scratch, which makes training deeper models easier. The mapping is not a guarantee of convergence, but it does reduce the chance of vanishing gradients and keeps the learned representation close to the original input. In practice, residual blocks that use identity shortcuts tend to train faster and reach higher accuracy on very deep networks. Thus, statements that identity mapping improves gradient flow, keeps representations stable, and is commonly used in residual architectures are true.