Exploding Gradient

  • weight matrices have max eigenvalue , gradient increases per layer
    • if enough depth, then converges to infinity