How to Shrink Large AI Models Using Knowledge Distillation
A method for teaching small, efficient AI models to mimic the complex decision-making patterns of much larger, more powerful neural networks.
Original patent title: “Training distilled machine learning models”
A method for teaching small, efficient AI models to mimic the complex decision-making patterns of much larger, more powerful neural networks. Granted to Google LLC in 2019 with 23 claims and 4 forward citations.
Key facts
Coverage
What does this patent actually cover?
This patent describes a process called knowledge distillation. First, a large, heavy 'cumbersome' model is trained on a dataset to learn complex patterns. Then, a smaller 'distilled' model is trained, not just to predict the correct answer, but to mimic the probability distribution (the 'soft outputs') of the large model. By using a 'temperature constant' higher than 1 during training, the model is forced to pay attention to the relationships between incorrect answers, which provides more information than a simple right-or-wrong label. This allows the smaller model to achieve performance levels close to the large model while being much faster and lighter for mobile devices.
The gap
What does this patent NOT cover?
- Does not cover training models from scratch without a pre-existing cumbersome model.
- Does not cover hardware-specific optimization techniques like model quantization or pruning.
- Does not cover methods where the distilled model is trained using only hard labels (e.g., just the correct class) instead of soft outputs.
- Does not cover architectures where the distilled model has more parameters than the cumbersome model.
These exclusions are unique to PatentBrief — derived from the actual claim language, not patent-office boilerplate.
What made this novel
The innovation is using a 'temperature' parameter to soften the output distribution, which reveals the 'dark knowledge'—the subtle hints about how the big model views the similarities between different categories.
Schematic visualization of the patent's claim structure. Hand-drawn diagrams in progress for each landmark patent.
Where you've seen this
Real-world examples
Mobile versions of Google Translate
On-device voice recognition on Android phones
Lightweight image classification models for mobile apps
Why it matters
The bigger picture
This technique is fundamental to modern AI deployment. It allows companies like Google to run sophisticated language models and image classifiers on smartphones and edge devices that lack the massive computing power required by the original, cumbersome models. It effectively bridges the gap between research-grade supercomputing and consumer-grade hardware.
Filed
June 4, 2015
Granted
May 14, 2019
Market context
Who's building on this
Companies in this space
Google remains a primary driver of this research, but the technique is now a standard practice across the AI industry. Companies like Meta, Microsoft, and various startups building on-device AI models use distillation to optimize their LLMs and computer vision systems for local execution.
Market impact
This patent formalized a practice that enabled the 'edge AI' revolution. By providing a reliable way to compress intelligence without losing significant accuracy, it allowed developers to move AI processing from expensive cloud servers directly onto consumer devices, reducing latency and improving user privacy.
Claim 1 — Plain English
What this patent covers
This patent describes a process called knowledge distillation. First, a large, heavy 'cumbersome' model is trained on a dataset to learn complex patterns. Then, a smaller 'distilled' model is trained, not just to predict the correct answer, but to mimic the probability distribution (the 'soft outputs') of the large model. By using a 'temperature constant' higher than 1 during training, the model is forced to pay attention to the relationships between incorrect answers, which provides more information than a simple right-or-wrong label. This allows the smaller model to achieve performance levels close to the large model while being much faster and lighter for mobile devices.
The clever bit
The innovation is using a 'temperature' parameter to soften the output distribution, which reveals the 'dark knowledge'—the subtle hints about how the big model views the similarities between different categories.
What it does not cover
- Does not cover training models from scratch without a pre-existing cumbersome model.
- Does not cover hardware-specific optimization techniques like model quantization or pruning.
- Does not cover methods where the distilled model is trained using only hard labels (e.g., just the correct class) instead of soft outputs.
- Does not cover architectures where the distilled model has more parameters than the cumbersome model.
Patent timeline
Application submitted to the patent office
Application published, typically 18 months after filing
Patent officially issued
PatentBrief Score
Impact Score
Moderate
Citation count
14/40
Early citations
Claim breadth
15/20
Broad claimsclaimsThe numbered statements at the end of a patent that legally define what the inventor owns.Read more →
Recency
10/20
Granted 5–10 years ago
Assignee scale
20/20
Major company or institution
PatentBrief Impact Score — based on citation count, claim breadth, recency, and assignee scale. Not a legal assessment.
Heuristic Value Estimate
What this patent might be worth
$52K – $166K
Midpoint $104K · 9.0 yr remaining · industry ×1.6
Heuristic only — blends forward/backward citation counts, claim scope, time remaining, litigation history, and CPC-derived industry baseline. Real valuations need a professional appraisal.
The original legal language
Original claims
23 claims as filed with the patent office.
Concepts involved
Citations
Patent lineage
Cite this patent
Vinyals, O., Hinton, G. E., & Dean, J. A. (2019). How to Shrink Large AI Models Using Knowledge Distillation (U.S. Patent No. 10,289,962). U.S. Patent and Trademark Office. https://patentbrief.org/patent/us/10289962/deep-q-networks-dqn
Auto-generated from the patent record. Double-check author order and the issue date against the official USPTO document before submitting.
Embed
Add this patent to your site
Drop this plain-English patent card into any blog post or article — free, no signup. It always links back to the full breakdown here.
<div data-patentlens-widget data-patent-number="US10289962"></div> <script src="https://patentbrief.org/embed.js" async></script>
Stay in the loop
Get a weekly digest of new patents.
One email per week. No spam. Unsubscribe anytime.
Keep exploring
Related patents you should know
US 4683195 · 1987
How to Make Billions of Copies of a DNA Segment
This patent describes the Polymerase Chain Reaction (PCR), a method to rapidly create many copies of a specific piece of DNA or RNA, enabling its detection and analysis.
Cetus Corp
US 8697359 · 2014
How to Edit Genes in Human Cells Using an Engineered CRISPR System
This patent describes an engineered CRISPR-Cas9 system for precisely cutting DNA in eukaryotic cells to change how genes work, opening the door for gene editing in complex organisms.
Massachusetts Institute of Technology
US 7657849 · 2010
How the iPhone's Slide-to-Unlock Gesture Works
Apple's 2010 patent describes unlocking a device by dragging a specific graphical image across the touchscreen along a predefined path, a gesture that became iconic with the original iPhone.
Apple Inc
US 4733665 · 1988
How Doctors Implant a Permanent Stent Using a Balloon
This patent describes the method for placing a permanent, expandable wire mesh tube inside a blood vessel or other body tube using a balloon-tipped catheter to widen it and keep it open.
Expandable Grafts Partnership
US 4965188 · 1990
How to Make Many Copies of a DNA Piece with Heat
This patent describes the Polymerase Chain Reaction (PCR) method, a technique to make millions of copies of a specific DNA segment using a heat-resistant enzyme and repeated temperature changes.
Cetus Corp
US 4235871 · 1980
How to Encapsulate Active Materials in Lipid Bubbles Efficiently
This patent describes a method for trapping biologically active substances inside tiny, multi-layered fat bubbles called liposomes, using a specific water-in-oil emulsion and gel-forming process to improve how much material gets captured.
Individual
More to explore
More in AI & Machine Learning
US 10452978 · 2019 · Google LLC
How AI Models Understand Language Using 'Attention'
US 6523026 · 2003 · Huntsman International LLC
How Computers Find Hidden Connections Between Different Fields of Knowledge
US 11615208 · 2023 · Capital One Services LLC
How Cloud Systems Automatically Create and Train AI Data Models
US 10402750 · 2019 · Facebook Inc
How Facebook Uses Deep Learning to Predict What You Might Like
New to patents?
Common Questions
Frequently Asked Questions
What does How to Shrink Large AI Models Using Knowledge Distillation cover?
A method for teaching small, efficient AI models to mimic the complex decision-making patterns of much larger, more powerful neural networks.
Who owns patent US 10289962?
Google LLC owns this patent, granted in 2019.
When does this patent expire?
This patent is expected to expire on May 14, 2039, when the invention enters the public domain.
What is patent US 10289962 cited by?
This patent has been cited by 4 later patents that build on its ideas.
What problem does this patent solve?
This technique is fundamental to modern AI deployment. It allows companies like Google to run sophisticated language models and image classifiers on smartphones and edge devices that lack the massive computing power required by the original, cumbersome models. It effectively bridges the gap between research-grade supercomputing and consumer-grade hardware.
What does this patent NOT cover?
Does not cover training models from scratch without a pre-existing cumbersome model.
Same assignee
More from Google LLC
Patent monitoring