Pruning
Removing weights from a trained model — setting them to zero — to reduce model size without retraining from scratch.
Think of it like
Editing a book by deleting unimportant sentences.
Example
A layer has 1 million weights. You remove 50% of them (the ones closest to zero) and the layer still works because those weights didn't contribute much.
How it actually works
Pruning reduces memory and compute proportionally. Structured pruning (removing whole neurons) is easier to implement but less efficient; unstructured pruning (removing individual weights) is more efficient but requires sparse matrix operations that aren't always fast on hardware. Modern approach: use sparsity-aware training (pruning during training) instead of post-hoc pruning.
For product teams
Cheaper than retraining but usually less effective than quantization.
For engineers
Start with unstructured magnitude pruning. If hardware doesn't support sparse ops, use structured pruning instead.
Related
- Compression technique.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome