A 30-layer network with sigmoid activations barely updates its early layers. What helps most?
Answer: B. ReLU-family activations, good initialization, normalization and residual connections
Sigmoid derivatives are at most 0.25, so gradients shrink layer by layer. Modern networks combine better activations, careful init, normalization and skip connections.