research
Noise Found the Connections Worth Keeping
A local fluctuation-based pruning rule preserved trained recurrent-network behavior better than magnitude pruning.
Summary
A local fluctuation-based pruning rule preserved trained recurrent-network behavior better than magnitude pruning.
Researchers tested noise-prune on task-trained recurrent neural networks and report that it preserved performance while removing connections, outperforming magnitude-only pruning and matching or exceeding a nonlocal second-order method. Sampling connections by estimated importance and rescaling those retained were both necessary, although the best rescaling was smaller than theory predicted. Biological plausibility here refers to locality of the rule, not proof that brains prune this way.
Why it matters
A local fluctuation-based pruning rule preserved trained recurrent-network behavior better than magnitude pruning.
Limits and context
- Researchers tested noise-prune on task-trained recurrent neural networks and report that it preserved performance while removing connections, outperforming magnitude-only pruning and matching or exceeding a nonlocal second-order method.
- Biological plausibility here refers to locality of the rule, not proof that brains prune this way.
Key claims
A local fluctuation-based pruning rule preserved trained recurrent-network behavior better than magnitude pruning.
Qualification: Researchers tested noise-prune on task-trained recurrent neural networks and report that it preserved performance while removing connections, outperforming magnitude-only pruning and matching or exceeding a nonlocal second-order method.
Evidence: source-2026-08-09-005
Sources
- arXiv preprint 2608.05464arXiv · primary research
Corrections
No corrections have been recorded for this story.