DirMixE:利用测试不可知长尾识别与层次标签变化等级
IEEE transactions on pattern analysis and machine intelligence
|December 23, 2025
概括
这项研究介绍了DirMixE,这是一种新的混合专家策略,用于测试无关的长尾识别. 它有效地解决了全球和本地标签分发的变化,改善了模型的概括性和稳定性.
科学领域:
- 计算机视觉 计算机视觉
- 机器学习 机器学习
- 人工智能的人工智能
背景情况:
- 由于未知和不平衡的测试标签分布,长尾识别任务存在挑战.
- 现有的混合专家 (MoE) 方法主要解决全球变化,忽视了当地分布转移.
研究的目的:
- 开发一种新的 MoE 策略 DirMixE,该策略将测试标签分布的全球和本地变化考虑在内,以改进长尾识别.
- 引入潜在技能微调 (LSF) 框架,用于对基础模型进行参数高效微调.
- 从理论上分析标准学习和参数有效微调 (PEFT) 的概括误差极限.
主要方法:
- 拟议的DirMixE策略将专家分配到迪里克莱特元分布中,以捕获本地变异,并固有全球变异处理.
- 开发了隐藏技能微调 (LSF) 框架,使用LoRA和适配器实现参数高效微调.
- 导出了概括错误的理论上限,并分析了PEFT假设类缩放.
主要成果:
- 在处理多样化和不平衡的测试标签分布方面,DirMixE证明了其有效性.
- 这种LSF框架使得基础模型能够高效地微调.
- 理论分析证实了基于差异的规范化的好处,并提供了对PEFT复杂性的见解.
结论:
- 通过解决分层标签分布变化,DirMixE为测试无关的长尾识别提供了一个强大的解决方案.
- LSF框架为参数有效的微调提供了一种有效的方法.
- 该研究为机器学习和PEFT的概括提供了理论理解.
更多相关视频
07:34Perceptual and Category Processing of the Uncanny Valley Hypothesis' Dimension of Human Likeness: Some Methodological Issues
Published on: June 3, 2013
17.8K
08:25Combining Eye-tracking Data with an Analysis of Video Content from Free-viewing a Video of a Walk in an Urban Park Environment
Published on: May 7, 2019
9.5K
相关概念视频
01:22Force Classification
2.2K
Forces play a crucial role in the study of physics and engineering. They are essential in describing the motion, behavior, and equilibrium of objects in the physical world. Forces can be classified based on their origin, type, and direction of action.
Contact and non-contact forces are two of the most widely used categories of forces. As the name suggests, contact forces require physical contact between two objects to act upon each other. Examples of contact forces include frictional,...
Contact and non-contact forces are two of the most widely used categories of forces. As the name suggests, contact forces require physical contact between two objects to act upon each other. Examples of contact forces include frictional,...
2.2K
03:31Labeling DNA Probes
9.2K
DNA probes are fragments of DNA labeled with a reporter tag to enable their detection or purification. The resulting labeled DNA probes can then hybridize to target nucleic acid sequences through complementary base-pairing, and may be used to recover or identify these regions.
Radioisotopes, fluorophores, or small molecule binding partners like biotin or digoxigenin, are the most widely used reporter tags for labeling DNA probes. These labels can be attached to the probe DNA molecule via...
Radioisotopes, fluorophores, or small molecule binding partners like biotin or digoxigenin, are the most widely used reporter tags for labeling DNA probes. These labels can be attached to the probe DNA molecule via...
9.2K