Home Glossary Vanishing Gradient

Vanishing Gradient

The vanishing gradient problem arises when gradients shrink repeatedly as backpropagation moves through many layers or time steps. Early layers then receive updates too small to learn useful long-range relationships, making training slow or ineffective. Saturating activation functions and repeated multiplication by values below one can worsen the effect, especially in older recurrent networks. Modern approaches reduce it through ReLU-like activations, careful initialization, residual connections, normalization, gated recurrent units, and architecture choices such as transformers. Monitoring gradient norms helps diagnose the issue. It differs from exploding gradients, where updates grow dangerously large, although both can appear in the same deep training system.

Related News