Demis Hassabis Patents & Innovation Profile — PatSnap Eureka
Demis Hassabis: Patent Portfolio & Innovation Analysis
Demis Hassabis is the co-founder and CEO of Google DeepMind, holding 23 patents assigned exclusively to DeepMind Technologies Limited, spanning neural network architectures, visual concept learning, and AI-driven algorithm optimisation for hardware devices, with filings across 8 jurisdictions from 2018 to 2025. His published research record of 122 papers — including the AlphaFold family with over 10,500 citations — constitutes a critical prior art corpus for any IP professional operating in the machine learning, computer vision, or AI-for-science sectors.
Patent Filing Activity
Peak year was 2018 with 11 filings, corresponding to the visual concept learning family filed across US, EP, WO, AT, and DE jurisdictions.
Demis Hassabis's Patent Filing Patterns
A two-phase filing strategy: an 11-patent burst in 2018 for visual concept learning, followed by a 2022–2023 acceleration for algorithm optimisation, mirroring DeepMind's published research milestones.
Annual Patent Filings
Peak year 2018 saw 11 filings across 5 jurisdictions from a single priority date; 2023 saw 7 filings in the algorithm optimisation family.
Technology Domain Breakdown
Neural Networks & Deep Learning (G06N3) accounts for 50% of the portfolio, with Computer Vision and Knowledge Representation as secondary clusters.
Core Areas of Innovation
Demis Hassabis's patent portfolio concentrates on two architecturally significant clusters: neural network-based visual concept learning and AI-driven algorithm optimisation for hardware devices.
Neural Networks & Deep Learning
10 patentsCovers the architecture and training of artificial neural networks with a focus on disentangled visual concept learning — enabling systems to map symbolic language inputs to generated visual outputs by decomposing scenes into disentangled latent representations. Also includes abstraction learning via hidden layer activation patterns.
- Learning visual concepts using neural networks (US20200234468A1)
- Learning visual concepts using neural networks (US11354823B2)
- Learning abstractions using patterns of activations of a neural network hidden layer (US12333435B1)
Knowledge Representation & Reasoning
4 patentsCovers machine reasoning and knowledge representation systems, specifically AI-guided search over candidate algorithm spaces using learned knowledge captured through tensor decompositions. Addresses the automation of algorithm discovery — replacing manual engineering with reinforcement learning-based search.
- Optimizing algorithms for hardware devices (WO2024074452A1)
- Optimizing algorithms for hardware devices (US20240127045A1)
Computer Vision
5 patentsFive patents under G06V10 and G06V20 address the generation and understanding of visual scenes guided by symbolic language inputs, relevant to image synthesis, visual question answering, and multimodal AI systems that interpret and produce visual content from linguistic prompts.
- Learning visual concepts using neural networks (WO2019011968A1)
- Learning visual concepts using neural networks (EP4053740A1)
Demis Hassabis's Highest-Impact IP
The 2018 "Learning visual concepts using neural networks" family dominates citations, with US20200234468A1 accumulating 19 forward citations — indicating foundational methodology in disentangled visual representation learning.
| Patent Number | Title | Year | Citations | Assignee | Status |
|---|---|---|---|---|---|
| US20200234468A1 | Learning visual concepts using neural networks | 2018 | 19 ↑ | DeepMind Technologies Limited | Active |
| US11354823B2 | Learning visual concepts using neural networks | 2018 | 7 ↑ | DeepMind Technologies Limited | Active |
| WO2019011968A1 | Learning visual concepts using neural networks | 2018 | 6 ↑ | DeepMind Technologies Limited | — |
| US20240127045A1 | Optimizing algorithms for hardware devices | 2022 | 4 ↑ | DeepMind Technologies Limited | Pending |
| US12333435B1 | Learning abstractions using patterns of activations of a neural network hidden layer | 2023 | 2 ↑ | DeepMind Technologies Limited | Active |
| WO2024074452A1 | Optimizing algorithms for hardware devices | 2023 | 1 ↑ | DeepMind Technologies Limited | — |
Demis Hassabis's Research Collaborators
Most Frequent Co-Inventors
Collaboration Highlights
Hassabis's patent collaborations fall into two distinct teams: the visual concept learning family is led by Alexander Lerchner and Christopher Paul Burgess, while the algorithm optimisation family centres on Matej Balog, David Silver, and Julian Schrittwieser from DeepMind's reinforcement learning group. The depth of co-authorship across both families indicates substantive technical involvement rather than supervisory naming.
- Alexander Lerchner 14 joint patents
- Matej Balog 12+ joint patents
- Christopher Paul Burgess 12 joint patents
- Irina Higgins 12 joint patents
- Matthew Botvinick 11 joint patents
Research Literature by Demis Hassabis
122 papers indexed · Research clusters across AI for structural biology (AlphaFold), AI-driven algorithm and mathematical discovery (AlphaTensor, AlphaDev), and neuroscience-inspired reinforcement learning.
| Title | Year | Citations | Venue / Source |
|---|---|---|---|
| Accurate structure prediction of biomolecular interactions with AlphaFold 3 | 2024 | 10,525 ↑ | Google DeepMind / Isomorphic Labs |
| AlphaFold Protein Structure Database | 2021 | 6,044 ↑ | DeepMind / EMBL-EBI |
| Protein complex prediction with AlphaFold-Multimer | 2021 | 2,643 ↑ | DeepMind |
| Improved protein structure prediction using potentials from deep learning | 2020 | 2,481 ↑ | DeepMind / Francis Crick Institute / UCL |
| AlphaFold Protein Structure Database in 2024 | 2023 | 1,599 ↑ | Google DeepMind / Seoul National University / EMBL |
AI for Structural Biology
The AlphaFold family (2020–2024) represents the most scientifically impactful body of work, with AlphaFold 3 alone accumulating over 10,500 citations since May 2024. These papers have reshaped drug discovery, protein engineering, and computational biology at a global scale.
AI-Driven Algorithm Discovery
The AlphaTensor (2022, 370 citations) and AlphaDev (2023, 127 citations) papers directly underpin the algorithm optimisation patents filed from 2022 onwards — one of the clearest examples of an inventor's literature directly foreshadowing their patent strategy.
Neuroscience-Inspired Reinforcement Learning
Earlier work on hippocampal memory, spatial cognition, and episodic memory forms the conceptual foundation for DeepMind's reinforcement learning architectures and memory-augmented neural network designs, including the "Reinforcement Learning, Fast and Slow" paper (2019, 423 citations).
Patent Jurisdictions
DeepMind's filing strategy for the Hassabis-named portfolio covers 8 jurisdictions, with the US and EP as primary markets and Taiwan specifically targeted for the hardware algorithm optimisation family.
Filing Markets
The US (7 patents) and EP (5 patents) serve as the primary commercial markets, while Taiwan's 3 patents are particularly telling — algorithm optimisation for hardware devices has been specifically filed in a jurisdiction dominated by TSMC, MediaTek, and other chipmakers. India's 2025 pending application signals expanding geographic ambition for the algorithm optimisation family.
Why Demis Hassabis's Portfolio Matters
Strategic implications for patent attorneys, in-house IP teams, and R&D strategists working in machine learning, computer vision, and AI for science.
FTO Considerations
The visual concept learning family (US11354823B2, US20200234468A1, EP4053740A1 and related grants) covers methods for training neural networks to disentangle visual concepts and generate images from symbolic inputs. Organisations developing multimodal AI systems, text-to-image models, or symbol-grounded visual reasoning systems should assess whether their architectures engage the claim scope of these active patents. The 2023 continuation US12333435B1 extends this family's effective life to at least 2043.
Prior Art Relevance
Hassabis's 122-paper research record constitutes an exceptional body of prior art. The most cited patent US20200234468A1 has accumulated 19 forward citations. Papers predating his patent applications — particularly neuroscience and early deep learning work from his UCL years — may be relevant to novelty and obviousness assessments against third-party patents in overlapping domains. The AlphaFold literature (10,525+ citations) is a critical reference for structural biology and drug discovery IP searches.
Demis Hassabis Patent Portfolio: Common Questions
Analyse Demis Hassabis's Full Patent Landscape
Search 208M+ patents and papers. Map the complete DeepMind portfolio, track prosecution status, and conduct FTO analysis — all in PatSnap Eureka IP.
References & External Sources
- USPTO — US20200234468A1: Learning visual concepts using neural networks
- USPTO — US11354823B2: Learning visual concepts using neural networks
- EPO Espacenet — WO2019011968A1: Learning visual concepts using neural networks
- WIPO — WO2024074452A1: Optimizing algorithms for hardware devices
- PatSnap — Global Innovation Intelligence Platform
- PatSnap Products — AI-Powered IP Tools
PatSnap Eureka searches 208M+ patents and papers to answer instantly.