MUMBAI, India, Sept. 22 -- Intellectual Property India has published a patent application (202641109316 A) filed by Amrita Vishwa Vidyapeetham on September 11, 2026, for System And Method For Pruning Neural Networks Using A Fermi-Dirac Gating Layer.
Inventors include Veerappan, Ravikumar Pandi; Dev, Raunak; Abraham, Reba Susan; Biju, Devaprabha S; and Sankar, Mydhily.
The application for the patent was published on September 18, 2026, under issue no. 38/2026.
Abstract: A system (100) and a method (400) for differentiable pruning of a neural network are disclosed. The system (100) receives a neural network including interconnected layers having trainable weights and associates each trainable weight with a generalized many-body Fermi-Dirac gating layer. The gating function determines, for each corresponding trainable weight, a probabilistic retention value as a function of combined effect of an energy associated with corresponding trainable weight, a learnable chemical potential, a self-energy correction, one or more interaction terms, one or more bias parameters, and one or more temperature parameters. The probabilistic retention value is applied to generate masked trainable weights for forward propagation and backward propagation while jointly optimizing trainable weights and gating parameters using gradient-based optimization. The disclosed system (100) and method (400) enable adaptive and correlated pruning, maintain differentiability during training, eliminate post-training pruning, reduce computational complexity, and support efficient deployment across multiple neural network architectures.
Disclaimer: Curated by HT Syndication.
