MUMBAI, India, July 7 -- Intellectual Property India has published a patent application (202621047996 A) filed by Manisha Satish Divate; and Krishna Santosh Pathak on April 15, 2026, for A System And Method For Regulating Effective Neural Network Capacity Via Sparsity-Induced Control And R-Sicr Generalization Bounding.
Inventors include Manisha Satish Divate; and Krishna Santosh Pathak.
The application for the patent was published on July 03, 2026, under issue no. 27/2026.
Abstract: Abstract; Title; "A System and Method for Regulating Effective Neural Network Capacity via Sparsity-Induced Control and R-SICR Generalization Bounding." In an aspect of the present invention a computer-implemented method and system for training neural networks by regulating effective capacity through sparsity-induced mechanisms are disclosed. The method introduces a framework known as Sparsity-Induced Capacity Regularization (SICR), which utilizes sparsity in neural activations as a dynamic control signal rather than a target for model compression. By computing effective capacity (C) as a product of parameter and activation norms, and inducing sparsity to reduce said activation norms, the system actively regulates the model's capacity during the training phase. A corresponding theoretical framework, the R-SICR bound, establishes a mathematical relationship between sparsity, capacity, and the number of training samples to provide provable generalization guarantees. This approach enables significant improvements in generalization performance across various neural architectures without reliance on parameter pruning or predefined sparsity targets, addressing the limitations of over parameterized system.
Disclaimer: Curated by HT Syndication.
