Deep Learning · Level 2 of 5
Vanishing Gradient
Gradients becoming too small to effectively update earlier computations.
Repeated multiplication by small derivatives can weaken the training signal.
Example
Early layers change very little in a deep sigmoid network.
Listen to the definition and example
Audio transcript
Vanishing Gradient. Gradients becoming too small to effectively update earlier computations. Repeated multiplication by small derivatives can weaken the training signal. For example: Early layers change very little in a deep sigmoid network.
Explore this concept
Why it matters
This helps explain why a neural model learns well, becomes unstable, or fails to generalize.
Start with
Quick recall question
Try answering before looking back at the definition.