Deep Learning Patents: Top Companies & Filing Trends 2026
- Concentrated but not cornered. the ranked leader holds 33,323 records, yet the top 5 combined are only 13.3% of all 494,414 records in scope — a long tail does most of the filing.
- Growth has cooled from its peak. filings ran from 3,042 in 2021 down to 1,672 in 2024, a 45% pullback across that span, even before the newest, still-incomplete years are counted.
- Momentum is slowing across every major filer. recent-year YoY changes for the leading assignees are all negative, from -43% to -87%, which says more about publication lag than about interest cooling.
Filing growth compares 2021 (3,042 records) with 2024 (1,672) — a three-year span. 2024 is the most recent year we treat as complete: publication lags filing by roughly 18 months, so 2025 onwards are still filling in and any growth rate that ends there would understate the field. Top-5 share is the combined record count of the five largest assignees divided by all 494,414 records in scope (CR5), not by the ranked leaders only.
What the deep learning patent record actually shows
Deep learning patenting spans a search space built around training datasets, model parameters, feature vectors and inference layered onto core neural-network and feature-extraction claims. That breadth is why the corpus runs to 494,414 published records: the terms sit inside almost every modern AI-adjacent filing, from vision pipelines to speech systems to healthcare informatics. The dataset covers filings published between 2015 and the 2026 data cut-off, with the most recent 12–18 months understated because publication lags filing.
Reading this landscape well means separating documents from families and separating peak-year filing counts from the newest, still-incomplete years. The figures below use records as the unit of scale but flag family-level ranking wherever the source data supports it, and the trend chart is read from 2021 backward for anything resembling a growth conclusion.
Filing trend and technology composition
Two views of the same corpus: how filing volume has moved year over year, and how records distribute across the IPC subclasses that define the technical sub-areas of deep learning.
Filing trend, 2017–2026
Filings rose from 1,060 in 2017 to a peak of 3,042 in 2021, then fell to 1,672 by 2024 — a 45% decline across that three-year window. 2025 and 2026 figures are shown for completeness but understate true filing activity because of publication lag.
IPC subclass composition
G06N (computing arrangements based on AI models) leads at 2.5% of all 494,414 records, followed by general digital data processing (G06F), image data processing (G06T), and image/video recognition (G06V), each near 1%. Speech and audio (G10L), healthcare informatics (G16H) and diagnostic/surgical applications (A61B) each sit below 0.5%, marking smaller but active application niches. Because records can carry multiple IPC codes, these shares sum to more than 100% and should not be added together.
Shares are the percentage of the 494,414 records in scope. A patent can carry several IPC classes, so the shares add up to more than 100%.
Go deeper on Deep Learning Patent Landscape with Eureka
This page is one run against one query. Ask Eureka your own question about deep learning patent landscape and every answer comes back with the patent numbers behind it.
Try EurekaA representative claim in the deep learning space
US11688200B2 — Joint facial feature extraction and facial image quality estimation using a deep neural network (DNN)
Systems and methods for joint feature extraction and quality prediction using a shared machine learning model backbone and a customized training dataset are provided. A computer system receives a training dataset of example images labeled with categories, then trains a deep neural network to jointly perform facial feature extraction and quality scoring for an input image using a common DNN backbone.Filed by Fortinet, Inc., published 2023-06-27 — illustrates the pattern of pairing a shared backbone with task-specific heads and a custom-labeled training set.
| # | Publication no. | Patent title | Citations |
|---|---|---|---|
| 1 | US20180075581A1 | Super resolution using a generative adversarial network | 863 |
| 2 | US5504675A | Method and apparatus for automatic selection and presentation of sales promotion programs | 634 |
| 3 | US20160174902A1 | Method and System for Anatomical Object Detection Using Marginal Space Deep Neural Networks | 531 |
| 4 | US20170270919A1 | Anchored speech detection and speech recognition | 521 |
| 5 | US20150100530A1 | Methods and apparatus for reinforcement learning | 520 |
| 6 | US5465308A | Pattern recognition system | 475 |
| 7 | US20150324690A1 | Deep Learning Training System | 468 |
| 8 | US20170337682A1 | Method and System for Image Registration Using an Intelligent Artificial Agent | 435 |
| 9 | US8527276B1 | Speech synthesis using deep neural networks | 432 |
| 10 | US20150066496A1 | Assignment of semantic labels to a sequence of words using neural network architectures | 426 |
Citation counts reflect influence within the searched corpus and skew toward older publications; they are a signal of technical reach, not of current commercial relevance.
Each row carries its publication number; clicking a row searches Eureka by that number.
Put your own technology through the same analysis
Eureka on the web
When you want the answer in the next five minutes.
The agent works the prompt against patents and technical literature, citing every source.
Run your analysis now →MCP server & REST API
When it has to run inside your own pipeline.
Patent search, landscape analysis and assignee resolution as MCP tools. Drop them into any agent framework, or call REST directly.
Browse MCP servers →What the filing data signals
Three read-outs from the corpus that matter for deciding where to file, partner or design around existing claims.
Leadership is real but not dominant
The leading assignee alone accounts for 33,323 records, and the top 10 combined reach 19.9% of all records in scope. That leaves roughly four-fifths of the corpus spread across a long tail of filers, including many single-filing entrants — a sign the core techniques have diffused well beyond a handful of labs.
Volume peaked in 2021 and has since pulled back
Filings rose steadily through the late 2010s, peaked at 3,042 in 2021, and fell to 1,672 by 2024. Because publication lag runs about 18 months, 2025-2026 figures are not yet reliable evidence of a continued slowdown — they simply have not finished arriving.
Core AI-model computing dominates the classification mix
G06N leads the IPC composition, ahead of general digital processing, image processing and image/video recognition, each near 1%. Smaller shares in speech, healthcare informatics and diagnostic applications show deep learning claims extending into regulated, application-specific domains rather than staying purely algorithmic.
Eureka can read the same corpus for gaps instead of for coverage: under-claimed branches adjacent to deep learning patent landscape, with the prior art for and against each one.
Who is filing, and where the openings sit
The ranked leaders combine large diversified electronics and semiconductor filers with a long tail of single- and few-filing entrants. Recent-year momentum has slowed across the board, which is consistent with publication lag rather than a genuine drop in R&D activity.
Diversified electronics filers hold the largest single positions
The top-ranked assignee's record count is roughly 4.4 times the fifth-place holder's 7,598, showing a steep drop after the very top of the ranking rather than a smooth gradient.
A competitive mid-tier sits just behind the top five
Tenth place still holds 5,478 records, and the top 10 combined reach 19.9% of all 494,414 records — evidence of several well-resourced programmes beyond the household names at the very top.
Every major filer shows a negative year-over-year figure
Recent-year momentum ranges from -43% to -87% across the leading assignees. Given that the newest publication years are structurally incomplete, this pattern should be read as a data-lag artifact, not as leaders retreating from deep learning.
| Assignee | Recent year | YoY |
|---|---|---|
| NVIDIA Corporation | 15 | -83% |
| Huawei Technologies Co., Ltd. | 13 | -50% |
| Intel Corporation | 4 | -87% |
| Microsoft Technology Licensing, LLC | 4 | -43% |
| Tencent Technology (Shenzhen) Co., Ltd. | 4 | -75% |
| Samsung Electronics Co., Ltd. | 3 | -82% |
| Qualcomm Incorporated | 3 | -70% |
| Google LLC | 1 | -83% |
Where to take this analysis
The dataset points to specific follow-up work depending on whether the goal is freedom-to-operate, partnership scouting, or spotting where to file next.
Map claims against the under-claimed branches
Speech-audio, healthcare informatics and diagnostic-surgical applications carry markedly lower record shares than core G06N computing claims, which is where a narrowly drafted claim is less likely to collide with dense prior art.
Explore white space in EurekaWatch the mid-tier, not just the leader
The gap between the top ranked assignee and fifth place is steep, but the mid-tier down to tenth place still holds thousands of records each — worth tracking for partnership or acquisition signal.
Track assignee momentum in EurekaTreat 2025-2026 filing counts as provisional
Publication lag means the last 18 months of filings are undercounted in any dataset pulled today; re-run the trend view periodically rather than drawing conclusions from the newest bars alone.
Set up a monitoring search in EurekaCommon questions about the deep learning patent landscape
The assignee ranking in this dataset is led by a single company holding 33,323 records, well ahead of the fifth-ranked filer at 7,598. However, the top 5 assignees combined account for only 13.3% of all 494,414 records in scope, and the top 10 combined reach 19.9% — so no single company or small group controls the field. A long tail of filers, many with only a handful of records, makes up the majority of the corpus.
Filing volume peaked in 2021 at 3,042 records and had fallen to 1,672 by 2024, a 45% decline over that span. That said, publication typically lags filing by around 18 months, so the 2025 and 2026 figures in any dataset pulled today are necessarily incomplete and should not be read as proof the field is cooling. The safest read is that filing activity has normalized after the 2021 peak rather than that it is collapsing.
Core AI-model computing, classified under IPC subclass G06N, is the largest single category at 2.5% of all 494,414 records. General digital data processing, image data processing, and image/video recognition each sit near 1%, while speech and audio analysis, healthcare informatics and diagnostic-surgical applications are all below 0.5%. Because a single record can carry multiple IPC classes, these percentages should be read individually rather than summed.
The lower-density IPC classes point to it: speech and audio processing, healthcare informatics, and diagnostic-surgical applications each represent well under 1% of the corpus compared with roughly 2.5% for core AI-model computing claims. That does not mean these areas are unpatented, but claim density is lower, which typically means more room for a narrowly drafted first claim without running into the dense prior art that surrounds core neural-network computing. Co-assignee collaboration patterns in the data also suggest cross-institution filings cluster in specific niches rather than spreading evenly.
US11688200B2, assigned to Fortinet, Inc. and published in 2023, claims a method of jointly training a deep neural network to perform facial feature extraction and image quality scoring from a single shared backbone, using a custom-labeled training dataset. It is a useful reference point for the pattern of combining multi-task heads on a common DNN backbone with a purpose-built labeled dataset, a structure that recurs across many records in this corpus. Anyone building a similar joint-task facial or image-quality pipeline should review its specific claim language for overlap before assuming clear freedom to operate.
Research Deep Learning Patent Landscape in depth with Eureka
Go past this page: query the whole deep learning patent landscape corpus yourself, in your own scope.
Every answer comes back with patent numbers you can open.
Disclaimer. This page is generated from Patsnap Eureka data drawn from a limited snapshot of global patent and scientific-literature records, and is provided for general information and reference only.
Patent data carries inherent limitations: recent filings (typically the most recent 18–24 months) are under-counted due to standard publication lag; counts may be reported at either a patent-family or a patent-record basis and are not always directly comparable; classification, applicant-name, and citation data may contain errors, duplicates, or omissions; and the underlying search query defines and constrains the scope shown. As a result, the analysis may be incomplete or inaccurate and may not reflect the full technology landscape.
Nothing on this page constitutes an exhaustive prior-art, novelty, freedom-to-operate, or validity search, nor does it constitute legal, financial, investment, or professional advice, and it should not be relied upon as such. Any patent, commercial, or strategic decision should be verified independently and reviewed with qualified patent, legal, and domain professionals. Patsnap makes no warranties, express or implied, as to the accuracy, completeness, or fitness for any particular purpose of the information presented.
Machine translation. Assignee and organisation names originally recorded in Chinese, Japanese or Korean have been rendered into English by an AI translation step so that the tables stay readable. These renderings are best-effort and may not match a company’s registered English name; the original name is what the underlying patent record carries, and it is what any Eureka query launched from this page uses.