/Patent Tracker
SamsungPatent applicationAI & softwareInterest 4/10

Samsung patents method to shrink AI models using fake images

Samsung describes a technique to compress large artificial intelligence models into smaller, more efficient versions without needing real data. The process uses a computer program to generate fake images that help a smaller model learn from a larger one by matching their internal calculations.

Why it mattersThis could help companies run complex AI software on devices with limited memory, such as phones or cars, by making the software smaller and faster.

In the filing’s own words

A method and system are provided. The method includes receiving, at a generator, a random input; producing, at the generator, one or more adversarial images based on the received random input; receiving, at a teacher network, the one or more adversarial images, wherein the teacher network includes a batch normalization layer comprising a mean and a variance of batch normalization layer inputs; receiving, at a student network, the one or more adversarial images; minimizing a maximum of a distance between a feature map of the teacher network and a feature map of the student network based on the one or more adversarial images; and constraining the generator based on a loss term defined according to a combination of a batch normalization statistics parameter derived from the batch normalization layer, an instance categorical entropy parameter, and a batch categorical entropy parameter calculated based on the one or more adversarial images generated by the generator, wherein the batch normalization statistics parameter is computed based on a statistical divergence between the mean and the variance of the batch normalization layer inputs of the teacher network and a corresponding statistic of the one or more adversarial images generated by the generator.

What exactly they claim (claim 1)

1. A method for model compression, comprising: receiving, at a generator, a random input; producing, at the generator, one or more adversarial images based on the received random input; receiving, at a teacher network, the one or more adversarial images, wherein the teacher network includes a batch normalization layer comprising a mean and a variance of batch normalization layer inputs; receiving, at a student network, the one or more adversarial images; minimizing a maximum of a distance between a feature map of the teacher network and a feature map of the student network based on the one or more adversarial images; and constraining the generator based on a loss term defined according to a combination of a batch normalization statistics parameter derived from the batch normalization layer, an instance categorical entropy parameter, and a batch categorical entropy parameter calculated based on the one or more adversarial images generated by the generator, wherein the batch normalization statistics parameter is computed based on a statistical divergence between the mean and the variance of the batch normalization layer inputs of the teacher network and a corresponding statistic of the one or more adversarial images generated by the generator.

Facts

Company
Samsung (Samsung Electronics Co., Ltd.)
Official title
METHOD AND APPARATUS FOR DATA-FREE NETWORK QUANTIZATION AND COMPRESSION WITH ADVERSARIAL KNOWLEDGE DISTILLATION
Publication number
US20260268148A1
Application number
19657781
Filed
Apr 24, 2026
Published
Sep 10, 2026
Inventors
4
Claims
20

Classified under

G06N 3/088 Physics & computingG06N 3/045 Physics & computing
Read the full filing (official PDF) ↗ Open in the Patent Tracker

More from Samsung in AI & software

Summaries are simplified from the filing’s own abstract and first claim. A patent filing is not a product announcement.
Patent data and drawings: U.S. Patent and Trademark Office (USPTO). Nebula is not affiliated with or endorsed by the USPTO.