Vanishing Gradient
Gradients shrinking toward zero as they propagate through many layers.
Overview
The vanishing gradient problem is that a gradient multiplied by many small derivatives decays toward zero before reaching the early layers, so they stop learning. It solves nothing; it is the reason very deep networks could not be trained for years. It exists as the problem that residual connections, gating and normalisation were each introduced to address.
Technology Graph
How Vanishing Gradient connects to the rest of the stack.
Uses and depends on
- Backpropagation technique Algorithms Vanishing Gradient based on Backpropagation High ✓ source
Applied to
- Layer Normalization concept Algorithms Layer Normalization solves Vanishing Gradient High ✓ source
- Residual Connection concept Algorithms Residual Connection solves Vanishing Gradient High ✓ source