Which of the following strategies most effectively addresses the degradation problem in deep networks while considering training efficiency, optimization techniques, and scalability issues?

The most effective strategy is to use residual connections that add identity shortcuts, allowing each block to learn a small residual change instead of a full mapping; this keeps gradients flowing and reduces the need for extremely deep layers. By inserting a few extra parameters per block, training remains efficient because the shortcut paths are inexpensive to compute and do not increase the number of trainable weights dramatically. Optimization techniques such as batch‑norm and ReLU are then applied to each block, further stabilizing gradients and speeding convergence. Because each block can be trained independently, the architecture scales easily to very deep networks without exploding memory usage. For example, a 100‑layer ResNet can converge in hours on a single GPU, whereas a plain network of the same depth would fail to learn due to vanishing gradients.