Intellectual Property India Publishes Patent Application for An Energy-Efficient Information Retrieval System Utilizing Quantized Neural Networks And Approximate-Nearest-Neighbor Search For Edge Deployment

MUMBAI, India, July 24 -- Intellectual Property India has published a patent application (202611059134 A) filed by Dr. Abdul Aleem; Dr. Krishna Kant Agrawal; Dr. Roshni Singh; Mr. Ataussamad; Dr. Harish Kumar Gr; Mr. Shamsu Zama Khan; Dr. Pooja Singh; and Dr. Mohd. Aquib Ansari on May 09, 2026, for An Energy-Efficient Information Retrieval System Utilizing Quantized Neural Networks And Approximate-Nearest-Neighbor Search For Edge Deployment.

Inventors include Dr. Abdul Aleem; Dr. Krishna Kant Agrawal; Dr. Roshni Singh; Mr. Ataussamad; Dr. Harish Kumar Gr; Mr. Shamsu Zama Khan; Dr. Pooja Singh; and Dr. Mohd. Aquib Ansari.

The application for the patent was published on July 17, 2026, under issue no. 29/2026.

Abstract: ABSTRACT TITLE: AN ENERGY-EFFICIENT INFORMATION RETRIEVAL SYSTEM UTILIZING QUANTIZED NEURAL NETWORKS AND APPROXIMATE-NEAREST-NEIGHBOR SEARCH FOR EDGE DEPLOYMENT The present invention relates to an energy-efficient information retrieval system (100) for edge deployment. The system (100) comprises an edge hardware platform (101) having a low-power processor (102), a non-volatile memory (103), and a sensor interface (104). A quantized neural network encoder (105) converts queries into compressed embedding vectors (106) using mixed-precision integer arithmetic with per-layer bit widths between 2 and 8 bits assigned offline by a layer-wise bit-width allocator (202) based on calibration sensitivity scores and stored as a static configuration table. An approximate nearest neighbor search engine (107) performs hierarchical indexed retrieval over a quantized vector database (108) using a hierarchical graph index (403), product quantization codebook store (402), and a hardware distance computation accelerator (404) implementing integer asymmetric distance evaluation. A power management unit (109) modulates clock frequency via a predictive controller using exponentially-weighted moving-average prediction with battery and thermal overrides. A communication interface (110) transmits results. [FIGURE 1]

Disclaimer: Curated by HT Syndication.