作者:Bingyu Liu, Zhen Zhao, Zhenpeng Li, Jianan Jiang, Yuhong Guo, Jieping Ye ; AI Tech, DiDi ChuXing
propose a feature transformation ensemble model with batch spectral regularization for the Crossdomain few-shot learning
Ensemble (Feature Transformation Ensemble Model)
就是对feature extraction network
提取出来的单个视觉特征进行多种正交变换(随机生成一个对称矩阵,并计算得到他的特征向量作为视觉特征的变换因子),计算预测的交叉熵损失BSR (Batch Spectral Regularization)
:这个技术最早是在NIPS2019的一篇文章[1]中提出的, 认为最下化特征矩阵(由一个batch的视觉特征组成)的奇异值有助于减少模型fine-tune过程中的负迁移影响。Feature Transformation Ensemble Model
和BSR
对模型带来的提升较多。[1] Chen X, Wang S, Fu B, et al. Catastrophic forgetting meets negative transfer: Batch spectral shrinkage for safe transfer learning[J]. 2019.