MUMBAI, India, Aug. 17 -- Intellectual Property India has published a patent application (202641093086 A) filed by Velammal Engineering College on July 31, 2026, for A Hardware-Accelerated Deep Learning Processing System For Real-Time Neural Network Training And Inf.
Inventors include Daya Florance D; Nisha P M C; Jeya Mohan H; and Kayalvizhi S.
The application for the patent was published on August 14, 2026, under issue no. 33/2026.
Abstract: The invention describes a sophisticated processing system which uses hardware acceleration to develop complex neural networks and can run a deep learning application at the edge also. By fusing a smooth multi-stage pipe-line the system controls the complete life cycle of an Al model that initiates with the massive gathering and refinement of raw data coming from different ambient and cloud sensors. After structuring the data, the system will offer an iterative design environment where architectural topologies are visualized and based on predicted performance refilled. At the center of the invention is the configuring of heterogeneous computing clusters, which utilize high-speed graphic processors and field programming gate arrays to speed up the training phase, which is the one that most consumes computational resources. In practical purposes, the system is endowed with an optimization engine that simplifies and prunes the neural architectures. This technique helps in decreasing the computation cost and memory usage while hardly compromising the accuracy of the model. Therefore, the resultant models can sit on small low-resource hardware. Once deployed, these optimized models allow for low-latency, real-time decision-making on edge devices such as self-driving drones and smart camera systems. Through strong feedback loop, the invention completes its cycle which collects operational field data for additional learning. This architecture allows for the continual return of the latest models to the deployment environment so that the intelligent system remains tuned to constantly changing the real word and operates most effectively and autonomously across a distributed network of devices.
Disclaimer: Curated by HT Syndication.
