MUMBAI, India, Nov. 2 -- Intellectual Property India has published a patent application (202611116590 A) filed by Sharda University on September 29, 2026, for System And Method For Preventing Model Collapse In Recursive Artificial Intelligence Training Using Real-World Anchoring.
Inventors include Dr. Neetu Faujdar; and Dr. Anjum Mohd Aslam.
The application for the patent was published on October 30, 2026, under issue no. 44/2026.
Abstract: A system (1000) and a method (2000) for preventing model collapse in recursive artificial intelligence training are disclosed. The system (1000) comprises a processing unit (100) and a storage unit (200) maintaining a real-world reference database (210) representing an independently verified real-world reference distribution. A synthetic data generation module (400) generates synthetic training data, a multi-layer synthetic data verification module (500) evaluates the synthetic training data, and an analysis module (600) determines entropy, diversity, and confidence measures. A distribution divergence monitoring module (700) determines a distribution divergence value using a Wasserstein distance, and a model-collapse detection engine (800) determines a collapse-risk level from a combination of the measures and the divergence value. An adaptive data-mixing controller (900) dynamically determines a real-to-synthetic data ratio for a subsequent training dataset, a recursive training module (910) trains a subsequent model, and an intervention controller (930) initiates corrective actions when the collapse-risk level exceeds a threshold, thereby anchoring recursive training to real-world data. Fig. 1
Disclaimer: Curated by HT Syndication.
