Intellectual Property India Publishes Patent Application for Neural Network Architecture With Surprise-Modulated Dual-Pathway Processing, Fixed-Size Retrieval Buffer, And External Associative Memory For Constant-Cost Sequence Modeling

MUMBAI, India, Aug. 31 -- Intellectual Property India has published a patent application (202621081661 A) filed by Suraj Tripathi on July 02, 2026, for Neural Network Architecture With Surprise-Modulated Dual-Pathway Processing, Fixed-Size Retrieval Buffer, And External Associative Memory For Constant-Cost Sequence Modeling.

Inventor includes Suraj Tripathi.

The application for the patent was published on August 28, 2026, under issue no. 35/2026.

Abstract: WINA: a neural network architecture replacing global self-attention with wave retention, surprise-weighted outer-product memory with parallelizable soft delta rule, and a fixed-size retrieval buffer of B entries for exact matching. A single surprise signal controls memory writes, pathway routing, and external memory management. The buffer operates over B fixed entries independent of sequence length N, achieving O(d²+Bd) inference cost. Validated at matched parameters with identical initialization on A100: superior recall and language modeling quality compared to matched-parameter GPT transformer.

Disclaimer: Curated by HT Syndication.