In the context of training deep neural networks, which of the following scenarios best illustrates the impact of the vanishing/exploding gradients problem on backpropagation, training stability, and the risk of overfitting?

The vanishing gradient scenario is best illustrated when a deep recurrent network trains on a long‑term sequence and the gradients shrink to near zero, so the earlier layers receive almost no updates; this makes backpropagation ineffective, training stalls, and the model tends to memorize short‑term patterns, increasing overfitting risk. In contrast, an exploding gradient situation causes huge weight updates that destabilize training, leading to wildly fluctuating loss values and preventing convergence, which also encourages the network to fit noise and overfit. Thus, the scenario where the loss plateaus after a few epochs while the gradient norms decay to zero demonstrates the vanishing gradient’s impact on stability and overfitting. A concrete example is training a 50‑layer CNN on image classification where the loss stops improving after the first epoch, indicating that gradients cannot propagate to earlier layers.