MUMBAI, India, July 24 -- Intellectual Property India has published a patent application (202641086034 A) filed by Indian Institute Of Technology Madras Iit Madras on July 14, 2026, for A System And Method For Configuring An Artificial Neural Network For Parameter-Efficient Processing.
Inventor includes Mahesh V Panchagnula.
The application for the patent was published on July 17, 2026, under issue no. 29/2026.
Abstract: A SYSTEM AND METHOD FOR CONFIGURING AN ARTIFICIAL NEURAL NETWORK FOR PARAMETER-EFFICIENT PROCESSING The present invention relates to a computing system (100) and method for resource-efficient neural network training. A processor (110) computes a loss curvature representation of a neural network using matrix-free techniques and decomposes it to form a weighted structural matrix. A pivoting factorization is applied to structurally rank the network's trainable parameters. Based on an adaptive threshold, the ranked parameters are dynamically partitioned into an active subspace and a secondary subset. An explicit affine prediction map is derived linking the active subspace to the secondary subset. During training, constrained backpropagation restricts gradient computations exclusively to the active subspace. Simultaneously, the affine prediction map deterministically updates the secondary subset without requiring gradient calculations. This significantly reduces memory footprint, optimizer state requirements, and communication bandwidth in distributed systems, while preserving the full architectural capacity of the network without permanently pruning parameters.
Disclaimer: Curated by HT Syndication.
