Exploding Gradient

Gradients (error signals) become extremely large during backpropagation, causing huge, unstable updates to network weights, leading to divergence, loss spikes, and model failure