MUMBAI, India, Oct. 5 -- Intellectual Property India has published a patent application (202641116286 A) filed by Chaitanya Bharathi Institute Of Technology Autonomous; and Chaitanya Bharathi Institute Of Technology Autonomous, Vidyanagar, Proddatur -; Andhra Pradesh, India on September 29, 2026, for System And Method For Dynamic Sparsity-Based Neural Network Optimization For Joint Accuracy- Energy Efficiency In Deep Learning Systems.
Inventors include Kodidala Venkata Sai Vandana; Muchumarri Santhamani; Kunapuli Bharathi; Lomada Ganga Bhavani; Mada Sai Meghana; Mayaluru Himai Vardhan Reddy; Kunapuli Naveen Kumar; Madaka Sivani; Nagururu Mohammad Arief; and Madduru Chandbee.
The application for the patent was published on October 02, 2026, under issue no. 40/2026.
Abstract: A system (100) and method decide, for each input, how much of a neural network is executed. An input complexity analysis module (104) computes a complexity score for the input data. A gating network (106) turns this score into a sparsity mask through a learnable threshold function, and a sparse pathway selector (108) runs only the layers, blocks or channels of the dynamic neural network layers (110) that the mask marks as active. An energy monitor unit (112) estimates the energy used, from a hardware power interface or from a count of executed operations, and an accuracy evaluation module (114) scores the output against labels during training or by prediction confidence during inference. An energy-accuracy optimizer (116) uses both measures to adjust the threshold within the resource budget of the deployment platform (120). Embodiments cover distillation from a dense teacher network, multimodal inputs and federated learning on edge devices.
Disclaimer: Curated by HT Syndication.
