MUMBAI, India, June 26 -- Intellectual Property India has published a patent application (202621052531 A) filed by Dr. Govind Rangnath Suryawanshi; Dr. Santoshkumar V. Chobe; Dr Swati Nikam; and Dr. Nalini S Jagtap on April 24, 2026, for An Intelligent Deep Learning Compression System For Energy-Constrained Devices Using Dynamic Weight Pruning And Low-Bit Quantization.
Inventors include Dr. Govind Rangnath Suryawanshi; Dr. Santoshkumar V. Chobe; Dr Swati Nikam; and Dr. Nalini S Jagtap.
The application for the patent was published on June 19, 2026, under issue no. 25/2026.
Abstract: Abstract An intelligent deep learning compression system is described in this invention as an appropriate candidate for application on energy constrained devices. Examples include embedded systems, IoT nodes, and mobile computing platforms. The compression achieved by use of this system relies on the dynamic pruning of weights and adaptive low bit quantisation to reduce the size of models, the computational complexity of processing these models, and the power consumed by performing inference with these models while achieving a high level of prediction accuracy. In contrast to conventional static compression techniques, this invention's compression system provides dynamic adjustment of model parameters at run time based upon real-time measurements of system conditions. An energy-aware controller collects and monitors various performance metrics such as energy utilized, processing load, memory required, and latency for carrying out an inference. Using this information, the controller will dynamically adjust the level of compression being applied to an individual model to maintain the optimal balance between performance and efficiency. The performance of each configuration is continuously monitored, and a feedback-driven optimization loop using reinforcement learning is employed allowing for the system to learn from previous decisions and improve the adaptability over time. This innovative deep learning compression system will provide a scalable and flexible approach for enabling the effective and efficient deployment of deep learning-based solutions in resource constrained environments resulting in faster inferences, lower energy consumption, and minimal degradation of accuracy. Ultimately, this invention will provide for the practical, sustainable implementation of sophisticated AI models to be deployed in real-world applications at the edge of the computing domain.
Disclaimer: Curated by HT Syndication.
