Backpropagation

Backlinks

• 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

• Vanishing Gradient
  

  Gradients used for weight updates during backpropagation, become extremely small as they travel back through layers, effectively halting learning in early layers