MUMBAI, India, Nov. 21 -- Intellectual Property India has published a patent application (202511076251 A) filed by Indian Institute Of Technology, Kanpur, Uttar Pradesh, on Aug 11, for 'system and method for training a deep neural network based on a posit-based mixed-precision framework.'
Inventor(s) include Vishesh Mishra; Mahendra Rathor; and Urbi Chatterjee.
The application for the patent was published on Nov. 21, under issue no. 47/2025.
According to the abstract released by the Intellectual Property India: "A system and method for training a deep neural network based on a posit-based mixed-precision framework is disclosed. The system comprises a posit input module, which receives one or more high bit-width posit numbers that include a sign bit, a regime field, an exponent field, and a fraction field, while a regime boundary detection module identifies the end of the regime field. An integrated posit compute module comprising a rounding hardware module determines the available bit-width for the fraction field, truncates the fraction field to that available bit-width, applies rounding logic, and generates a lower bit-width posit number. A posit multiply and accumulate module performs a multiply and accumulate operation on the lower bit-width posit numbers to generate one or more weighted posit values, while a training module is configured to train the deep neural network based on the generated weighted posit values."
Disclaimer: Curated by HT Syndication.
