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 →Filing growth compares 2021 (81 records) with 2024 (107) — 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 667 records in scope (CR5), not by the ranked leaders only.
Knowledge distillation covers a specific family of claims: a smaller student model trained against the outputs, logits or intermediate features of a larger teacher, typically softened by a temperature parameter or matched at the feature level. The search underlying this page combines those mechanism terms with a second layer — temperature scaling, soft labels, capacity gap, data-free distillation — to separate genuine distillation art from adjacent model-compression filings that only mention teacher-student training in passing.
667 records fall in scope across the 2015-2026 window this dataset covers, concentrated heavily in computing-based-AI-model classifications but reaching into image recognition, speech processing, and a smaller set of applied domains such as healthcare informatics. The filing curve rises through the last complete year, and the assignee ranking shows one clear leader ahead of a broad, moderately active field.
Two views of the same 667 records: the year-by-year filing curve, and the IPC subclasses those filings claim against.
Filings ran from 2 in 2017 to a peak of 157 in 2025, with 40 already recorded for 2026. Because publication lags filing by roughly 18 months, the 2025 and 2026 figures will still grow as later filings publish; 2021 to 2024 is the most recent span that can be read as a complete trend, and over it filings rose 32%, from 81 to 107.
G06N (computing based on AI models) covers 80.5% of the 667 records, with G06F and G06V each above a quarter. Applied-domain classes are much thinner: G10L (speech/audio) sits at 8.1%, G06Q (business/commerce) at 5.7%, and G16H (healthcare informatics) at just 4.2% — the technique is broadly claimed in the abstract, far less so once tied to a specific application vertical.
Shares are the percentage of the 667 records in scope. A patent can carry several IPC classes, so the shares add up to more than 100%.
This page is one run against one query. Ask Eureka your own question about knowledge distillation techniques and every answer comes back with the patent numbers behind it.
Try EurekaFiled by Huzhou University, this record applies knowledge distillation to a TSK fuzzy classifier: the fuzzy rule structure combines an antecedent built on soft-label information from a distillation-temperature-adjusted output distribution with a consequent based on the original data. It is a concrete example of distillation mechanics migrating from pure neural architectures into hybrid neuro-fuzzy classifiers.Published 2023-12-14.


| # | Publication no. | Patent title | Citations |
|---|---|---|---|
| 1 | US20180268292A1 | Learning efficient object detection models with knowledge distillation | 484 |
| 2 | US20180158552A1 | Interpretable deep learning framework for mining and predictive modeling of health care data | 250 |
| 3 | US20190205748A1 | Soft label generation for knowledge distillation | 162 |
| 4 | US20200302295A1 | System and method for knowledge distillation between neural networks | 140 |
| 5 | CN110674880A | 用于知识蒸馏的网络训练方法、装置、介质与电子设备 | 109 |
| 6 | CN111027060A | 基于知识蒸馏的神经网络黑盒攻击型防御方法 | 93 |
| 7 | CN114241282A | 一种基于知识蒸馏的边缘设备场景识别方法及装置 | 87 |
| 8 | US20210117948A1 | Mobile device platform for automated visual retail product recognition | 87 |
| 9 | US20220344049A1 | Decentralized artificial intelligence (AI)/machine learning training system | 86 |
| 10 | US20220094713A1 | Malicious message detection | 69 |
Citation counts favour older filings simply because they have had more time to accumulate citers; read them as a signal of influence within this corpus, not as a ranking of current importance.
Patent titles are shown in the language they were filed in, not translated, so that each record stays verifiable against the original filing — a translated title will not match in Eureka or in any national register. Each row carries its publication number; clicking a row searches Eureka by that number.
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 →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 →Three patterns worth acting on before drafting new claims in this space.
The leading assignee holds 44 records; by fifth place the count is down to 12, and by tenth to 9. That gap between first and the rest suggests one organisation has built a genuinely broad portfolio while the remaining activity is comparatively dispersed across many single-digit filers.
Filings rose from 81 in 2021 to 107 in 2024 — the last span that can be read as complete once the ~18-month publication lag is accounted for. Treat 2025's 157 and 2026's 40 as undercounts still filling in, not as a peak already passed.
The overwhelming majority of filings claim distillation as a general AI-model technique (G06N) rather than tying it to a specific vertical. Healthcare informatics (4.2%) and business/commerce (5.7%) are comparatively open, suggesting domain-specific distillation claims face less prior art than generic ones.
China's receiving office accounts for the largest single share of records at 277, ahead of the United States at 193 and WIPO's PCT route at 54. Europe, India and the UK each show more modest but active filing counts, indicating this remains a genuinely global filing pattern rather than one concentrated in a single jurisdiction.
Eureka can read the same corpus for gaps instead of for coverage: under-claimed branches adjacent to knowledge distillation techniques, with the prior art for and against each one.
The ranked leaders span large technology companies, chipmakers and a handful of academic institutions; recent-year momentum data shows several of the most historically active filers with no recorded filings in the latest year, which is likely more a publication-lag artefact than an actual halt.
The top-ranked assignee's 44 records stand well clear of the fifth-place count of 12, giving it a portfolio breadth no other single filer in this dataset currently matches.
Vellore Institute of Technology recorded 4 filings in the latest year, down 33% year-on-year — the strongest recent-year momentum in the ranking, which is itself a modest figure against a field where several historically large filers show zero filings in the same year.
Only 10 co-assignee pairs appear across the dataset. The strongest, between Disney Enterprises and ETH Zurich, reaches 8 shared records — a rare example of sustained industry-academic co-filing in a field otherwise dominated by single-assignee filings.
| Assignee | Recent year | YoY |
|---|---|---|
| VELLORE INSITUTE OF TECH | 4 | -33% |
| International Business Machines Corporation (IBM) | 0 | — |
| Huawei Technologies Co., Ltd. | 0 | — |
| Samsung Electronics Co., Ltd. | 0 | -100% |
| Tencent Technology (Shenzhen) Co., Ltd. | 0 | -100% |
| Qualcomm Incorporated | 0 | -100% |
| NVIDIA Corporation | 0 | -100% |
| Presagen Pty Ltd | 0 | — |
The dataset points to specific next steps depending on whether you are scoping a filing, tracking a competitor, or evaluating freedom to operate.
Healthcare informatics, speech/audio and data-free distillation all carry thinner IPC shares than the core G06N claims. A quick prior-art check there is cheaper than assuming the whole field is occupied.
Explore white space in EurekaA 44-record leader with no recorded latest-year filings is more likely a publication-lag gap than a real slowdown. Track filing dates against publication dates before concluding activity has stopped.
Set up assignee tracking in EurekaThe highest-cited records in this dataset define the language much of the later art responds to. Understanding what they actually claim narrows the search for a defensible angle.
Review key patents in EurekaThis dataset identifies 667 published records in scope between 2015 and mid-2026, using a search that combines core distillation mechanism terms (knowledge distillation, teacher-student training, logit matching) with technique-specific terms like temperature scaling, soft labels and capacity gap. That figure covers a defined search string rather than every patent that ever mentions distillation in passing, so it should be read as a bounded, reproducible corpus rather than an absolute count. Family-level counting was used for the assignee ranking to avoid inflating totals through continuations or multi-jurisdiction refiling of the same invention.
The ranking covers 100 assignees, with one filer clearly ahead of the rest at 44 records against a fifth-place count of 12 and a tenth-place count of 9. The filers behind the leader include large technology companies, chipmakers, and academic and research institutions, with the top 5 combined holding 16.6% of the 667 records in scope and the top 10 combined holding 24.1%. That leaves the majority of filings spread across a long tail of assignees with only a handful of records each.
Yes, through the last year that can be read as a complete trend. Filings rose from 81 in 2021 to 107 in 2024, a 32% increase over that span. Figures for 2025 and 2026 look lower or flat by comparison, but that is a publication-lag artefact — patent applications typically publish around 18 months after filing, so the most recent one to two years will always understate true filing activity until later records catch up.
The overwhelming majority, 80.5% of the 667 records, fall under G06N (computing arrangements based on specific computational models, i.e. AI models), reflecting distillation's origin as a general model-compression technique. G06F (electric digital data processing) and G06V (image/video recognition) each cover just over a quarter of records, with G06T (image data processing), G06K (data recognition) and G10L (speech/audio) trailing behind. Applied-domain classes like G16H (healthcare informatics) and G06Q (business/commerce) are the thinnest, at 4.2% and 5.7% respectively, marking them as comparatively open relative to the core technique.
The clearest gaps sit in domain-specific applications rather than the core mechanism: healthcare informatics (4.2% of records), speech and audio processing (8.1%), and business/commerce applications (5.7%) all carry far thinner claim density than the general AI-model classification they build on. Data-free distillation and capacity-gap-aware teacher selection are also mentioned in the search terms but do not dominate the composition data, suggesting they remain less densely claimed sub-mechanisms. Anyone drafting in these areas should still run a targeted prior-art search, since thin composition shares indicate room rather than guarantee absence of blocking claims.
Go past this page: query the whole knowledge distillation techniques 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.
Help us improve
Something wrong, out of date, or missing on this page? Tell us and it goes straight to the team that maintains it.
This page