4 captured new canonical articles plus 2 cross-listings from the Noam Shazeer section. Lead author of the foundational Transformer patent (US 10,452,978, cross-listed). First inventor on the local-self-attention computer vision patent (US 20210390410, HaloNet) that scales attention to long sequences and images. Co-inventor on MultiModel (US 20200089755 — the unified multi-task multi-modal architecture that is the structural ancestor of MUM and Gemini), the fully-attentional computer vision system (US 20250292560, the ViT lineage), and the fast-decoding-with-discrete-latents method (EP 3732627) that makes LLM inference economically viable at search-scale latency. Spans 2017 to 2025.
About the Ashish Vaswani, Google Transformer, Attention-Based Vision & Fast Decoding Patents track
4 captured new canonical articles plus 2 cross-listings from the Noam Shazeer section. Lead author of the foundational Transformer patent (US 10,452,978, cross-listed). First inventor on the local-self-attention computer vision patent (US 20210390410, HaloNet) that scales attention to long sequences and images. Co-inventor on MultiModel (US 20200089755 — the unified multi-task multi-modal architecture that is the structural ancestor of MUM and Gemini), the fully-attentional computer vision system (US 20250292560, the ViT lineage), and the fast-decoding-with-discrete-latents method (EP 3732627) that makes LLM inference economically viable at search-scale latency. Spans 2017 to 2025.
Foundational Transformer & Multi-Modal Architecture
- Attention-Based Sequence Transduction Neural Networks (Transformer) (US 10,452,978 · October 22, 2019)
- Multi-Task Multi-Modal Machine Learning System (MultiModel) (US 20200089755 · March 19, 2020)
- Attention-Based Image Generation Neural Networks (US 12,142,034 · November 12, 2024)
Attention-Based Computer Vision
- Local Self-Attention Based Computer Vision Neural Networks (HaloNet) (US 20210390410 · December 16, 2021)
- Fully Attentional Computer Vision (Stand-Alone Self-Attention) (US 20250292560 · September 18, 2025)
LLM Serving & Decoding Efficiency
- Fast Decoding in Sequence Models Using Discrete Latent Variables (Latent Transformer) (EP 3732627 · November 4, 2020)
Why this inventor matters
Each inventor track inside the Nizam SEO War Room patents archive isolates one engineer's research arc — typically a decade or more of continuations, divisionals, and follow-up patents on a coherent research thread. Reading by inventor (rather than by topic) recovers the narrative: how the original disclosure evolved, what the continuations added, which claims got carved out into divisional applications, and how the thread eventually intersected with other research lines at Google or Microsoft. This is how working SEOs build durable intuition about search-engine internals — not by memorizing claim language, but by following the research bibliography that shipped the algorithms we now optimize against.
How to read this track
Start with the earliest filing — it sets the foundational disclosure. Continuations refine the claims; divisional applications split out separable inventions; the follow-up patents tend to introduce performance optimizations, edge-case handling, or downstream integration with other systems. Each patent on this site is annotated with the ranking surface it touches — query understanding, document retrieval, ranking, behavioral signals, knowledge graph, or AI search — so the practitioner can map the research back to the algorithm output observed on live SERPs.