Simplify Transfer Learning (简化迁移学习）

Transfer Learning

DNN has shown the very significant performance in different areas. But, it is not easy to leverage it in your cases. When leveraging DNN to solve your real problems, you are always challenged by the following hard problems.

1 Lack of labeled data

As you known, training a DNN is hard. So many weights need to be adjusted in a DNN, for training a DNN normally needs huge amount of training samples, such as, millions of training samples.In real world, you might not have enough labeled data to train a DNN, and labeling data is always expensive.

2 No Bible for constructing DNN

Deep learning is still in its early stage. Although, some typical DNN structures have been proved being successful in some areas, it is still no mature theory or methodology to decide the detail structure of DNN. For example, we know CNN (Conventional Neural Network) is a good choice for computer vision problems, but there is not mature theory or methodology on deciding the detail structure of the CNN, Such as, the number of layers. It is a time consuming task to experiment the different DNN structures.

Transfer learning is to address the above problem. Transfer learning is about reusing a successful DNN in similar cases. With transfer learnin