MUMBAI, India, Aug. 24 -- Intellectual Property India has published a patent application (202641098400 A) filed by Easwari Engineering College; and Srm Institute Of Science And Technology,ramapuram Campus on August 14, 2026, for System And Method For Predicting Neuron Importance In Large Language Models Using Stratified Activation-Patching Labels And Surrogate Regression For Efficient Model Pruning.
Inventors include Manish Varma D; Nelash U; and Dr. G. Jeyaram.
The application for the patent was published on August 21, 2026, under issue no. 34/2026.
Abstract: The present invention relates to a computer-implemented system and method for predicting neuron importance in a trained neural language model and pruning the said model. A feature extraction layer computes for every neuron, without ablation, an output-projection norm, an input-projection norm, a layer index, a mean absolute gradient and a sparsity value. A sampling layer partitions the neurons into gradient strata and draws a measurement set constituting a proper subset. A measurement layer labels only the said subset by mean-ablation activation patching, overwriting each selected neuron with its own mean activation and accumulating the resulting drop in probability of a correct continuation across a confidence-filtered multi-task probe set. A surrogate regression model is trained upon the labelled pairs, validated by rank correlation, and applied to all remaining neurons to predict importance without further causal measurement. A pruning layer removes a configured lowest fraction and reports retained quality per task category.
Disclaimer: Curated by HT Syndication.
