Residual Connections, and a Regularization Plan for a Small Dataset

Part 1. An interviewer asks: "A 56-layer plain CNN has worse training error than a 20-layer one. Is that overfitting? Why do residual connections fix it?"

Part 2. You then get a concrete task: classify 20 plant diseases from 5,000 labeled photos. A ResNet trained from scratch reaches 99% training accuracy and 62% accuracy on a 1,000-image validation set.

1. Answer Part 1 with the gradient of a residual block, and a numerical illustration of vanishing in a plain stack.
2. Propose a prioritized regularization plan for Part 2 with specific values, and explain how you would tell which pieces actually help. (Compute the standard error of the 62% estimate.)