BriefGPT.xyz
Jun, 2023
神经网络变异性崩溃的定量化
Quantifying the Variability Collapse of Neural Networks
HTML
PDF
Jing Xu, Haoxiong Liu
TL;DR
本文提出了一种名为VCI的新指标，用于量化神经网络中可变性崩溃现象，该指标与线性探测损失密切相关，具有不变性和数值稳定性，并证实其可以表征预训练神经网络的可转移性。
Abstract
Recent studies empirically demonstrate the positive relationship between the
transferability
of
neural networks
and the within-class variation of the last layer features. The recently discovered
→