1、 What is the difference between transfer learning and fine tuning?

One 、 The migration study

for instance , hypothesis boss Let you do some target detection , This data set is about optical fiber box inkjet character detection . The problem is , Data sets are few （ Only 1000 Data sheet ）, There is much disturbing information in it , You find that starting from zero training yolo The effect is very poor , It's easy to over fit . What shall I do? , So you think of using Transfer Learning, Use what others have already trained Imagenet To make a model of .

Two 、 What are the methods of transfer learning

1. hold Alexnet Inner convolution Take out the features of the output of the last layer , And use it directly SVM classification . This is a Transfer Learning, Because you used Alexnet Have learned in “ knowledge ”.
2. Vggnet Take out the final output of convolution layer , Classify with Bayesian classifier . The thought is basically the same .
3. Until you can put Alexnet、Vggnet The output of is taken out and combined , Design a classifier to classify by yourself . In this process, you not only use Alexnet Of “ knowledge ”, Also used. Vggnet Of “ knowledge ”.
4. Last , You can also use it directly fine-tune This method , stay Alexnet On the basis of , Add the full connection layer again , Then train the network .

3、 ... and 、fine-tune Use policy

There are many factors that determine how to use transfer learning , This is the most important, only two ： The size of the new dataset 、 And the similarity between the new data and the original data set . One thing to remember ： The first few layers of the web learn about common features , The next few layers learn about the features related to categories . Here are four scenarios to use ：

1. The new data set is small and similar to the original data set . Because the new data set is smaller , If fine-tune It may be over fitting ; And because old and new data sets are similar , We expect them to have similar high-level characteristics , We can use the pre training network as a feature extractor , Train the linear classifier with the extracted features .
2. The new data set is similar to the original data set . Because the new data set is big enough , Sure fine-tune The whole network .
3. ** The new dataset is small and not similar to the original dataset .** The new data set is small , It's better not to fine-tune, Unlike the original data set , It's better not to use high-level features . At this time, we can use the features of the front layer to train SVM classifier .
4. ** The new data set is large and not similar to the original data set .** Because the new data set is big enough , You can retrain . But in practice fine-tune The pre training model is useful . The new data set is big enough , Sure fine-tine The whole network .

Four 、 summary

Sum up ,Transfer Learning The concern is ： What is? “ knowledge ” And how to better use what we got before “ knowledge ”. There can be many ways and means . and fine-tune Just one of the means .