一种基于数据压缩的显式-隐式混合数据集蒸馏系统及方法

By using an explicit-implicit hybrid dataset distillation system to jointly optimize the bitrate-utility loss, the problem of insufficient bitrate optimization in existing technologies is solved, achieving efficient dataset representation and performance improvement.

CN120768373BActive Publication Date: 2026-07-17HARBIN INSTITUTE OF TECHNOLOGY (SHENZHEN) (INSTITUTE OF SCIENCE AND TECHNOLOGY INNOVATION HARBIN INSTITUTE OF TECHNOLOGY SHENZHEN)
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202510858638.7
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2025-06-25
Publication Date
2026-07-17
Estimated Expiration
2045-06-25

AI Technical Summary

Technical Problem

Existing dataset distillation techniques fail to effectively consider the impact of bit-level bit rate on storage and transmission, resulting in an unsatisfactory optimization between storage costs and training utility.

Method used

An explicit-implicit hybrid dataset distillation system is adopted. By jointly optimizing the rate-utility loss, and utilizing explicit latent representation modules, implicit generator modules, autoregressive modules, and entropy coding calculation modules, efficient representation and parameterization of datasets are achieved.

Benefits of technology

While reducing storage and transmission requirements, it maximizes the retention of useful information from model training, improves the overall performance of dataset distillation techniques, and demonstrates excellent bitrate-utility performance.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120768373B_ABST
    Figure CN120768373B_ABST
Patent Text Reader

Abstract

本发明公开了一种基于数据压缩的显式‑隐式混合数据集蒸馏系统及方法，方法包括：通过显式潜在表示模块表征原始数据集图像的多分辨率层次信息，生成若干不同分辨率的潜在表示；将潜在表示输入隐式生成器模块，生成合成数据集，并计算合成数据集与真实数据集的蒸馏损失；通过自回归模块估计潜在表示的概率分布，计算潜在表示的码率大小；基于熵编码计算模块计算显式潜在表示、隐式生成器模块参数及自回归模块参数的码率损失；将码率损失与蒸馏损失进行联合优化，更新显式潜在表示模块、隐式生成器模块及自回归模块的参数，输出优化后的合成数据集及合成数据集对应的码率；本发明实现了馏损失和码率损失的联合优化。
Need to check novelty before this filing date? Find Prior Art

Citation Information

Patent Citations

  • Image compression quality enhancement method based on knowledge distillation

    CN112929663A

  • Data set distillation and weight reduction method based on deep utilization of GAN prior enhancement

    CN118378078A