MUMBAI, India, Aug. 12 -- Intellectual Property India has published a patent application (202621075931 A) filed by Symbiosis International Deemed University on June 17, 2026, for System And Method For Loss Landscape Analysis In Over-Parameterized Neural Networks For Safety-Critical Perception.
Inventors include Dr. Akhil Gupta; Ayush Bisen; Sai Konde; and Madhura Rewatkar.
The application for the patent was published on August 07, 2026, under issue no. 32/2026.
Abstract: ABSTRACT SYSTEM AND METHOD FOR LOSS LANDSCAPE ANALYSIS IN OVER-PARAMETERIZED NEURAL NETWORKS FOR SAFETY-CRITICAL PERCEPTION The present invention provides a system and method (100) for geometric and spectral analysis of loss landscapes in over-parameterized neural networks for safety-critical perception models. The system comprises a synthetic data generation module (110), a model scaling and over-parameterization control module (120), a neural network training engine (130) employing a Sharpness-Guided Adaptive Training algorithm (150), a loss landscape geometry analysis module (140) for computing basin width and loss surface variance, a perturbation and stability evaluation module (145) for computing loss sensitivity and adversarial robustness, a Hessian spectral analysis module (160) for computing dominant eigenvalues spectral gap and effective rank using deflated power iteration, and a failure-direction alignment module (170) for correlating failure patterns with dominant curvature directions. The unified output integrates geometric, spectral, perturbation and failure alignment metrics for comparative safety assessment across model architectures. [
Disclaimer: Curated by HT Syndication.
