Deep Learning · Level 2 of 5
Gradient Clipping
Limiting gradient values or their norm before an update.
Clipping controls unusually large updates without fixing every cause of instability.
Listen to the definition and example
Audio transcript
Gradient Clipping. Limiting gradient values or their norm before an update. Clipping controls unusually large updates without fixing every cause of instability. For example: A trainer caps the total gradient norm at one.
Explore this concept
Why it matters
This helps explain why a neural model learns well, becomes unstable, or fails to generalize.
Quick recall question
Try answering before looking back at the definition.