MUMBAI, India, June 22 -- Intellectual Property India has published a patent application (202641068730 A) filed by Renuka P; S. Ashok Kumar; Jayamoorthy S; Jm Hamer Shield; S Rajalakshmi; Buvana Jeeva; Abirami K; Dr. K. Nirmala; Chetna Aakash Upasani; and Dr. Nisha Thakur on June 01, 2026, for Energy-Efficient Deep Learning Model Compression Method.
Inventors include Renuka P; S. Ashok Kumar; Jayamoorthy S; Jm Hamer Shield; S Rajalakshmi; Buvana Jeeva; Abirami K; Dr. K. Nirmala; Chetna Aakash Upasani; and Dr. Nisha Thakur.
The application for the patent was published on June 12, 2026, under issue no. 24/2026.
Abstract: Abstract The current invention describes an energy efficient method of deep learning model compression ischemic energy efficient method of deep learning model compression ischemic examine deep learning application for mobile, embedded, and resource constrained devices, like IoT nodes, wearables, and smartphones. This compression technique, for the first time, we introduced a full interface to multi-step compression that includes mixed precision quantization adaptive to model weight structure. We help guide knowledge distance, and compression to help decrease the model footprint and the operating costs significantly and dynamically. In addition, we employ a multi-cell on-the-fly real-time monitoring system and engine that allows a feedback-creating system to modulate device energy states based on real-time device energy states and feedback sensors, to modulate the operating cost of the adaptive scale model of the device compression. This technique enables the deep learning model to be applied in real-time, low-power AI system environments. Thus, the invention improves the dependability of the model, the AI performance under the compression conditions, the battery life, and the overall system reduces the operating system and the overall system performance. Along with it given the efficiency of the system and the compression system. Reduce the overall compressions for Time instabilities. In a system, we provide the efficiency elevation and the elevation, we enable Time instabilities overall, and provide the efficiency overall for time instabilities used the system for time. The proposed methodology follows a logical, stepwise approach starting with pre-analysis of the base model. Then, we focus on Hybrid Compression, along with, stepwise Real-Time Deployment tuning for optimization. Each step generates performance metrics that influence the subsequent stage to enable smart model tuning on the full lifecycle of the model. The method guarantees that despite a significant cut in computational work, the reduced model maintains the proper knowledge topology required to predict accurately. It also features a self-optimizing design, adaptable from feedback from the device, thus applicable in many different real-time settings.
Disclaimer: Curated by HT Syndication.
