Yoshua Bengio Patents & Innovation Profile — PatSnap Eureka
Yoshua Bengio: Patent Portfolio & Innovation Analysis
Yoshua Bengio is a Canadian computer scientist and Turing Award laureate who holds 36 patents across 11 jurisdictions, spanning neural network data processing systems, dialect-adaptive speech recognition, and multimodal deep learning architectures, with filings from 1997 to 2021. His portfolio is primarily assigned to Imagia Cybernetics Inc. and reflects a selective but strategically significant engagement with formal IP in two distinct commercial phases — first at AT&T in the late 1990s, then through applied AI ventures from 2017 onward.
Patent Filing Activity
Peak year was 2017 with 10 filings, all through Imagia Cybernetics Inc., followed by 9 filings in 2019 through the Samsung collaboration.
Yoshua Bengio's Patent Filing Patterns
Bengio's filing history shows two distinct phases separated by a 19-year gap: early AT&T filings in 1997–1998, then a concentrated burst of applied AI patents from 2017 onward.
Annual Patent Filings
2017 was the peak year with 10 filings through Imagia Cybernetics, followed by 9 Samsung speech recognition filings in 2019.
Technology Domain Breakdown
Neural network data processing (G06F15) is the dominant domain with 10 patents, closely followed by speech recognition (G10L15) with 9 patents.
Yoshua Bengio's Core Areas of Innovation
Bengio's patent portfolio concentrates in neural network data processing and speech recognition, with additional coverage in network architectures, compression, and document recognition.
Neural Network Data Processing Systems
10 patentsThis dominant domain covers architectures where neural network modules process structured data including graphs, feature maps, and multimodal inputs. Patents range from the foundational Graph Transformer Network paradigm to modern multimodal processing systems that remain robust when input modalities are missing — a critical challenge in medical imaging and multi-sensor inference.
- Module for constructing trainable modular network in which each module inputs and outputs data structured as a graph (US6128606A)
- Method and system for processing a task with robustness to missing input information (WO2017158575A1)
- Method and system for processing a task with robustness to missing input information (US20190073563A1)
Dialect-Adaptive Speech Recognition
9 patentsThis cluster covers a speech recognition architecture co-developed with Samsung Electronics and Université de Montréal. The patents disclose a method for generating dialect-specific parameters from a parameter generation model, applied dynamically to a trained acoustic model to produce a dialect-adapted recogniser — addressing a longstanding robustness problem where monolithic ASR models degrade on non-standard dialects.
- Speech recognition method and apparatus (US20200126534A1)
- Speech recognition method and apparatus (EP3640934A1)
- Device and method for recognising voice, and device and method for training voice recognition model (JP2020067658A)
Neural Network Architectures
2 patentsThis subset addresses neural network training procedures and foundational deep learning architecture methods. These patents relate to the theoretical underpinnings of gradient-based learning and network design that informed much of Bengio's subsequent applied work.
- Neural network training architecture methods
- Gradient-based learning procedures for modular network systems
Adaptive Binary Arithmetic Coding
2 patentsThe Z-coder patents, co-invented with Léon Bottou at AT&T, cover a fast adaptive binary arithmetic coder designed for improved probability estimation and decoding speed. This infrastructure-level compression technology is relevant to signal transmission and data encoding systems, and was developed during Bengio's period at AT&T Bell Labs.
- Z-coder adaptive binary arithmetic coding system (CA2244380A1)
- Z-coder adaptive binary arithmetic coding system (CA2244380C)
Document & Pattern Recognition
1 patentThis domain covers the application of trainable modular network architectures to document recognition tasks, including optical character recognition and structured document parsing. The Graph Transformer Network approach introduced in the G06F15 cluster has direct application in document analysis workflows, reflected in this IPC classification.
- Graph Transformer Network applied to document recognition
- Modular trainable network for structured document parsing
Yoshua Bengio's Highest-Impact IP
US6128606A leads with 99 forward citations — the 1997 Graph Transformer Network patent co-invented with Léon Bottou and Yann LeCun at AT&T remains the most cited work in the portfolio by a significant margin.
| Patent Number | Title | Year | Citations | Assignee | Status |
|---|---|---|---|---|---|
| US6128606A | Module for constructing trainable modular network in which each module inputs and outputs data structured as a graph | 1997 | 99 ↑ | AT&T Corporation | Inactive |
| WO1998040824A1 | Module for constructing trainable modular network in which each module inputs and outputs data structured as a graph | 1998 | 14 ↑ | AT&T Corp. | Inactive |
| WO2017158575A1 | Method and system for processing a task with robustness to missing input information | 2017 | 12 ↑ | Imagia Cybernetics Inc. | — |
| US20200126534A1 | Speech recognition method and apparatus | 2019 | 12 ↑ | Samsung | Active |
| US20190073563A1 | Method and system for processing a task with robustness to missing input information | 2017 | 9 ↑ | Imagia Cybernetics Inc. | Inactive |
| CN109313627A | Method and system for processing a task with robustness to missing input information | 2017 | 3 ↑ | 映佳控制公司 | Inactive |
| CA3017697A1 | Method and system for processing a task with robustness to missing input information | 2017 | 2 ↑ | Imagia Cybernetics Inc. | Active |
| EP3640934A1 | Speech recognition method and apparatus | 2019 | 2 ↑ | Samsung | Active |
Yoshua Bengio's Research Collaborators
Most Frequent Co-Inventors
Collaboration Highlights
Bengio's patent collaboration network divides cleanly along the two phases of his filing history. The Imagia Cybernetics cluster is built around a tight four-person team — Nicolas Chapados, Nicolas Guizard, and Mohammad Havaei each appear in 11 joint filings, suggesting a deliberate research and commercialisation partnership in medical imaging AI. The Samsung collaboration introduced a different team led by Yoo Sang Hyun with 7 joint filings, reflecting the co-development agreement between Université de Montréal and Samsung's speech technology division. The early AT&T phase involved Léon Bottou and Yann LeCun, both of whom became equally prominent figures in AI research.
- Nicolas Chapados 11 joint patents
- Nicolas Guizard 11 joint patents
- Mohammad Havaei 11 joint patents
- Yoo Sang Hyun 7 joint patents
Research Literature by Yoshua Bengio
939 papers indexed · Research spans sequence-to-sequence learning, encoder-decoder architectures, neural machine translation, and deep learning fundamentals
| Title | Year | Citations | Venue / Source |
|---|---|---|---|
| Learning Phrase Representations using RNN Encoder–Decoder for Statistical Machine Translation | 2014 | 13,335 ↑ | Université de Montréal / Ecole Polytechnique de Montréal |
| On the Properties of Neural Machine Translation: Encoder–Decoder Approaches | 2014 | 4,952 ↑ | Université de Montréal |
| Graph Attention Networks | 2018 | 2,117 ↑ | Université de Montréal / Universitat Politècnica de Catalunya |
| Practical Recommendations for Gradient-Based Training of Deep Architectures | 2012 | 1,154 ↑ | Université de Montréal |
| On Using Very Large Target Vocabulary for Neural Machine Translation | 2015 | 710 ↑ | Université de Montréal |
Sequence-to-Sequence Learning & Neural Machine Translation
The largest research cluster covers encoder-decoder architectures and neural machine translation. The 2014 RNN Encoder-Decoder paper (13,335 citations) introduced the foundational architecture for neural translation systems and the attention mechanism era, with direct influence on subsequent transformer-based models.
Speech Recognition & Acoustic Modelling
A substantial body of work covers end-to-end acoustic modelling, convolutional networks for ASR, gated recurrent units, and spoken language understanding — research that directly informed the Samsung speech recognition patents filed from 2018 onward and the dialect-adaptive architecture disclosed in those filings.
Deep Learning Fundamentals & Representation Theory
The third cluster addresses training dynamics, optimisation methods, biologically plausible learning algorithms, and representation theory. The 2012 gradient-based training recommendations paper (1,154 citations) remains a key reference for practitioners designing deep architectures, and constitutes prior art relative to many subsequent patent filings in the field.
Patent Jurisdictions
Bengio's 36 patents span 11 jurisdictions, with the United States, European Patent Office, and Canada receiving the most filings, reflecting the commercial priorities of AT&T, Imagia Cybernetics, and Samsung.
Filing Markets
The jurisdictional distribution reflects a commercially rational prosecution strategy: broad US and EP coverage for the core inventions, with selective extension into Asia-Pacific markets where the assignees had active commercial operations or competitive exposure. The Singapore and Hong Kong filings reflect Imagia Cybernetics' multimodal processing family, while Japan and South Korea coverage corresponds to Samsung's home and key export markets for speech technology products.
Why Yoshua Bengio's Portfolio Matters
Strategic implications for patent attorneys, in-house IP teams, and R&D strategists working in AI, speech technology, and medical imaging.
FTO Considerations
The multimodal feature map processing family — centred on the "robustness to missing input information" architecture — covers a design pattern with broad applicability in medical AI, autonomous systems, and multi-sensor applications where input dropout is a realistic operating condition. Several family members expired in 2024–2025, entering the public domain. However, active family members in the US and Canada (including CA3017697C, granted January 2021) remain live, and claim scope should be assessed carefully before designing around this architecture. The Samsung speech recognition family is fully active across US, EP, CN, JP, and KR, and covers a dialect-parameter generation approach that any commercial ASR developer working on accent or dialect adaptation should review.
Prior Art Relevance
Bengio's academic record is arguably more significant than the patent portfolio for anyone filing in deep learning-adjacent domains. His 2014 publications on encoder-decoder architectures and the broader body of work on recurrent networks, attention mechanisms, and sequence modelling — including the RNN Encoder-Decoder paper with 13,335 citations — form a dense layer of prior art that predates many subsequent patent applications by other filers. IP teams conducting patentability assessments in NLP, speech, or sequence learning should include Bengio's 939-paper academic output — not just his patents — in their search strategy. The patents appear to formalise specific implementations of concepts validated through prior academic publication, making publication dates critical for prior art dating.
Yoshua Bengio Patent Portfolio — Common Questions
Analyse Yoshua Bengio's Full Patent Portfolio
Access citation mapping, claim-level search, FTO analysis, and inventor network intelligence across all 36 patents and 939 papers in PatSnap Eureka IP.
References & External Sources
- US6128606A — Graph Transformer Networks, AT&T Corporation (1997). View on Google Patents · View on PatSnap Eureka
- WO2017158575A1 — Robustness to missing input information, Imagia Cybernetics Inc. (2017). View on WIPO PatentScope
- EP3640934A1 — Speech recognition method and apparatus, Samsung (2019). View on EPO Espacenet
- Mila — Quebec Artificial Intelligence Institute. mila.quebec
- PatSnap Eureka IP — AI-native patent intelligence platform. patsnap.com
PatSnap Eureka searches 208M+ patents and papers to answer instantly.