Machine Learning 务实----Applying deep learning to real-world problems

1. Pre-tuning method 

 在现实世界里应用ML，得到大量精确标注的数据是昂贵的。

  如果只有少量精确标注的数据，pre-tuning method可以帮助提升最后训练模型的精度[1]。

  First pre-tuning on cheap large datasets on related domain.

  Then fine-tuning on expensive well-labeled data.

 As we fine-tune on precisely labeled data, 

 it is possible to pre-train on so-called weakly labeled data.

 (i.e. 90% of the labels might be correct and 10% wrong)

2.Caveats of real-world label distributions[1]

 在现实世界里，得到的数据有以下特征（相对于学术界里的平衡数据）：

 - Unbalanced label distribution

 - Unbalanced cost of misclassification

 Solution[1],让训练集里的数据数量上更平衡:
- More data
- Change labeling:ie. merge some rare labels
- Sampling: ignore / over-or undersample / Negative minging / weighting the loss

3. Understanding black box models
在现实世界里应用ML，你不仅需要提高准确率，还需要考虑以下问题：
- understand why and how a model can make wrong predictions,
- give some intuition why our model can perform better than any previous solution,

- make sure that the model cannot be tricked.  

 

参考： 

[1] 《Applying deep learning to real-world problems》

   https://medium.com/merantix/applying-deep-learning-to-real-world-problems-ba2d86ac5837