迁移学习(Transfer learning)
1 Introduction
- Example:Dog/Cat Classifier
- 现实生活中一直存在迁移学习
2 Classification
label & label: {(xs,ys),(xt,yt)}
- Fine tuning:复制前面几层layer的实验效果,复制越多越差
- multitask learning:共用前面几层layer;共用中间的layer;
- 不同语种之间的迁移
- Progressive NN:迁移其他任务训练得到的神经网络
label & unlabel: {(xs,ys),(xt)}
- Domian-adversarial training:
- feature exteactor → Domian classifier → domain label 骗过
- feature exteactor → label predictor 保证
- 《Unsupervised Domain Adaptation by Backpropagation》——ICML2015
- 《Domain-Adversarial Training of Neural Networks》——JMLR2016
- Zero-shot learning:不同的任务,例如source data是猫和狗,test data是草泥马
- 只预测得到接近的类,接近的特征
- attribute embedding:利用神经网络等模型映射到映射空间中,越接近的点越像
- f∗,g∗=argminf,g∑nmax(0,k−f(xn)g(yn)+maxm≠nf(xn)g(ym)) ,k由自己定义
- Domian-adversarial training:
unlabel & label:self-taught learning
- unlabel & unlabel:self-taught clustering