MUMBAI, India, June 22 -- Intellectual Property India has published a patent application (202621044129 A) filed by Dr. Anjali Devi Patil; Dr. Sanjay M. Patil; Prof. Anil Satyadeo Londhe; Mr. George Yashraj Janier; Mr. Vedant Vikas Javir; Ms. Siddhi Hemant Kambli; and Ms. Mansi Sachin Khanvilkar on April 07, 2026, for "adaptive Multi-Phase Efficient Training System And Method For High-Fidelity Generative Adversarial Networks'.
Inventors include Dr. Anjali Devi Patil; Dr. Sanjay. M. Patil; Prof. Anil Satyadeo Londhe; Mr. George Yashraj Janier; Mr. Vedant Vikas Javir; Ms. Siddhi Hemant Kambli; and Ms. Mansi Sachin Khanvilkar.
The application for the patent was published on June 12, 2026, under issue no. 24/2026.
Abstract: The present invention relates to a system and method for efficient training of generative adversarial networks (GANs) using an adaptive multi-phase framework. The system comprises a generator, a discriminator, a data loader, and an adaptive training controller configured to dynamically regulate training operations. The method includes distillation-based initialization, wherein semantic features from a pre-trained model are transferred to the generator using a mapping network and feature projector. The system further employs adaptive sparse training, wherein gradient-based importance scoring and loss ratio monitoring are used to dynamically adjust sparsity and optimize computational efficiency. Additionally, a one-shot multi-resolution training mechanism enables simultaneous generation and evaluation of outputs at multiple resolutions using a shared backbone and hierarchical discriminator. The adaptive training controller dynamically allocates computational resources and maintains training stability. The invention reduces training time, improves resource utilization, and enhances generative performance, making it suitable for deployment in resource-constrained and high-performance computing environments.
Disclaimer: Curated by HT Syndication.
