The difference between transfer learning and finetuning

For example, suppose today's boss gives you a new dataset that allows you to sort the images, and this dataset is about flowers. The problem is that there are very few flower in the dataset, and there is not much data in the dataset, and you find that the effect of training CNN from zero training is very poor and easy to fit. What to do, so you think of using transfer Learning, with other people have trained good imagenet model. There are many ways to do this:
The characteristics of the last layer output of the convolution layer in the alexnet are taken out and then directly classified by SVM. This is transfer Learning, because you use the "knowledge"that Alexnet has learned. The final output of the Vggnet convolution layer is taken out and classified by Bayesian classifier. Thought basically ditto. You can even combine the output of alexnet and vggnet to design a classifier. In this process you not only use the alexnet "knowledge", but also use the vggnet "knowledge". Finally, you can also use the fine-tune method directly, on the basis of alexnet, add the full connection layer, then to train the network.
In summary, Transfer learning is concerned about what is "knowledge" and how to make better use of the "knowledge" previously obtained. There are many ways and means to do this. And fine-tune is just one of the means.

Author: the source of any
Link: https://www.zhihu.com/question/49534423/answer/127022241
Source: Know
Copyright belongs to the author. Commercial reprint please contact the author to obtain authorization, non-commercial reprint please indicate the source.