New AI pruning method preserves RNN performance with noise

Researchers Sanjith Senthil and Rishidev Chaudhuri have introduced **noise-prune**, a novel unsupervised pruning rule for recurrent neural networks (RNNs) that leverages noisy fluctuations to evaluate connection importance. Unlike traditional methods that rely solely on connection magnitude or require computationally expensive second-order information, noise-prune uses stochastic sampling to preserve connections while rescaling retained synapses to maintain overall network strength.

In their arXiv paper (arXiv:2608.05464), the team demonstrates that noise-prune maintains task performance in functional RNNs, outperforming simple magnitude-based pruning by a significant margin. The method’s key innovation lies in its biologically plausible approach—sampling connections probabilistically based on their noise-derived importance scores and rescaling retained connections to preserve average synaptic strength. While theoretical predictions suggested higher rescaling factors, empirical testing revealed optimal performance at lower rescaling levels, highlighting the method’s adaptability in practical scenarios.