Residual connections, introduced in ResNet architectures, allow gradients to flow through networks without vanishing by adding the input of a layer to its output. This technique enables the training of much deeper neural networks by effectively addressing the degradation problem associated with increasing depth.