MUMBAI, India, July 7 -- Intellectual Property India has published a patent application (202641061734 A) filed by Suresh S; M. Sulthan Alavudeen; Dr. R. Sujatha; Dr. N, Thinaharan; Dr, K. Santhoshkumar; Dr. B. Chitradevi; and Dr, B, Gunasundari on May 15, 2026, for Neural Latent Space Compression System For Cross Model Knowledge Transfer.
Inventors include Suresh S; M. Sulthan Alavudeen; Dr. R. Sujatha; Dr. N, Thinaharan; Dr, K. Santhoshkumar; Dr. B. Chitradevi; and Dr, B, Gunasundari.
The application for the patent was published on July 03, 2026, under issue no. 27/2026.
Abstract: ABSTRACT A novel Neural Latent Space Compression System for Cross-Model Knowledge Transfer is proposed to enable efficient transfer of knowledge between heterogeneous deep learning models with reduced computational complexity and memory usage. The framework employs a Dual-Stage Variational Latent Compression Network (DS-VLCN) that converts high-dimensional mentor network features into compact latent representations for apprentice network learning. The system integrates Multi-Scale Feature Alignment (MSFA) for semantic feature mapping, a Variational Autoencoder-based Latent Compression Algorithm (VAE-LCA) for efficient encoding, and Graph-Guided Knowledge Distillation (GGKD) using Graph Neural Networks to preserve structural relationships during transfer. A Contrastive Latent Consistency Loss (CLCL) further enhances feature discrimination and transfer accuracy. The proposed system shows improved compression efficiency, faster inference, and higher knowledge transfer performance compared with conventional distillation methods, making the framework suitable for edge Al, federated learning, and resource- constrained intelligent systems
Disclaimer: Curated by HT Syndication.
