MUMBAI, India, May 1 -- Intellectual Property India has published a patent application (202641049200 A) filed by Srm Institute Of Science And Technology, Ramapuram Campus; and Easwari Engineering College, Chennai, Tamil Nadu, on April 17, for 'adaptive multi-step transformer-diffusion interaction (mtdi) model.'
Inventor(s) include Dr. D. Rajalakshmi; Ms. G. Sumathi; Dr. M. S. Minu; Ms. V. Gowri; S Eswaran; V B Shreesha; and Nilesh Siva SV.
The application for the patent was published on May 1, under issue no. 18/2026.
According to the abstract released by the Intellectual Property India: "Multivariate Time-Series Forecasting (MTSF) is an important part of various domains, including financial market analysis, power grid load management, weather analysis and industrial IoT applications. While Hybrid Transformer-Diffusion models have achieved state-of-the-art accuracy, their iterative sampling process introduces latency and increased computational cost. We propose Adaptive-MTDI, which optimizes the Multi-Step Transformer-Diffusion Interaction (MTDI) model through three key innovations: 1) A Dynamic Contextualization module that predicts a Complexity Score to modulate denoising effort based on the volatility of the sequence; 2) A Parallel initialization method for ensemble sampling; and 3) An Adaptive DPM- Solver++ implementation with Mixed-Precision Caching. Our model reduces inference latency by over 45% and Video RAM (VRAM) consumption by 39% compared to the base MTDI model. These optimizations enable the deployment of hybrid transformer-diffusion forecasting in real-time environments without much hardware constraints."
Disclaimer: Curated by HT Syndication.
