few-shot learning(三):Pretraining and Fine Tuning

优点:简单,准确度高。

数学基础

余弦相似度

few-shot learning(三):Pretraining and Fine Tuningfew-shot learning(三):Pretraining and Fine Tuning

softmax

few-shot learning(三):Pretraining and Fine Tuning

 softmax会把大的值变大,小的值变小。

few-shot learning(三):Pretraining and Fine Tuning

  Few-shot Prediction Using Pretrained CNN

预训练一个大网络提取特征,预测时用到这个网络。可以使用多种方法训练(训练方法以及网络结构对结果会有影响),但是使用监督学习训练完,要把全连接层去掉。

few-shot learning(三):Pretraining and Fine Tuning

 few-shot分类方法

拿Query与μ做对比。

few-shot learning(三):Pretraining and Fine Tuning

 few-shot learning(三):Pretraining and Fine Tuningfew-shot learning(三):Pretraining and Fine Tuning

Fine-Tuning

Fine-Tuning可以大幅提高准确率。之前的讨论,W与b是固定值。

few-shot learning(三):Pretraining and Fine Tuning

 其实,可以在Support Set上微调这两个参数。

few-shot learning(三):Pretraining and Fine Tuningfew-shot learning(三):Pretraining and Fine Tuning

考虑防止过拟合:

few-shot learning(三):Pretraining and Fine Tuning

Trick1: 好的初始化

few-shot learning(三):Pretraining and Fine Tuning

 Trick2: 正则化

few-shot learning(三):Pretraining and Fine Tuningfew-shot learning(三):Pretraining and Fine Tuning

 Trick3: 余弦相似度+SoftMax分类器

few-shot learning(三):Pretraining and Fine Tuningfew-shot learning(三):Pretraining and Fine Tuning

总结

few-shot learning(三):Pretraining and Fine Tuning