Deep Learning Network

Since there are already really advanced network, why not use transfer learning on most of the projects? Why should I create my own network for my own implementation? Are there any benefits of training my own network with my own data?

It depends on a type of problem, fund, time, and individual choice. There is no one answer.

A large model train on classifying human faces cannot perform good to classify complex medical images or something. But a large model train on classifying complex medical images can be used for another similar task, with little tunning.


In addition to @saifkhanengr‘s great reply:

Transfer Learning works well:

  • in general if you have (implicit) representative features learned in the pre-trained model which help in the fine-tuning
  • e.g. if you learn / transform a certain style, see also: transfer of style

In the end a deep learning model can learn what is contained in the data. If the data is sufficiently representative, compared to the new specialised tasks, transfer learning should be worth a look! Then you can leverage the pre-trained model (which is usually also high efforts, super expensive and difficult to create in the first place).

see also this thread.

Best regards

1 Like