MUMBAI, India, July 24 -- Intellectual Property India has published a patent application (202641086440 A) filed by Vellore Institute Of Technology on July 15, 2026, for “a Method For Generating An Optimized Convolutional Neural Netw Ork Model”.
Inventors include Dr. Swetha. N. G.; Ayushi Jha; Pritam Satpathy; and Manya Singhal.
The application for the patent was published on July 17, 2026, under issue no. 29/2026.
Abstract: The present invention relates to a method and system for generating an optimized convolutional neural network (CNN) model are disclosed. The method includes training a baseline CNN to obtain initial weights and performance metrics, computing composite importance scores for filters across selected convolutional layers, and progressively pruning channels based on the computed scores. The pruned model is dynamically rebuilt by reconstructing affected layers and restoring inter-layer compatibility. A differentiable neural architecture search is then performed on the rebuilt model using a shared multi-objective efficiency score that jointly optimizes predictive accuracy, computational cost, parameter count, and inference latency, enabling convergence toward a Pareto-optimal architecture. The optimized model is subsequently fine-tuned using adaptive learning strategies to recover performance. The pruning and architecture search processes operate in an interdependent iterative pipeline with a feedback loop, ensuring coordinated optimization. The system comprises modules for data processing, training, pruning, rebuilding, architecture search, evaluation, and deployment, and supports export of the optimized compact model into multiple formats including ONNX, TFLite, and TorchScript for deployment across diverse computing platforms such as edge devices, mobile systems, and cloud environments. Fig 1 to 2.
Disclaimer: Curated by HT Syndication.
