MUMBAI, India, Sept. 28 -- Intellectual Property India has published a patent application (202511132712 A) filed by Indian Institute Of Technology, Mandi on December 27, 2025, for Method And System For Training Neural Network Models.
Inventors include Dr. Hitika Tiwari; and Dr. Satyajit Thakor.
The application for the patent was published on September 25, 2026, under issue no. 39/2026.
Abstract: A system (100) and a method (700) for training a neural network are disclosed. A batch of training data including input feature vectors and associated target values is received and model outputs are generated by applying the neural network model to the input feature vectors. A training loss is computed from a comparison between the model outputs and the target values. Per-example gradient information describing sensitivity of the loss to model parameters is determined and used to derive kernel information characterizing relationships between training examples. A spectral analysis of the kernel information yields kernel modes, including a set of top modes. Residual errors are projected onto the top modes to obtain mode coefficients. An adaptive update direction that combines a baseline gradient direction and a mode- based direction is formed, and model parameters are updated according to an underlying optimization routine. FIG. 1
Disclaimer: Curated by HT Syndication.
