Transfer Learning is one of the popular topics in machine learning, more specifically deep learning where the knowledge of a trained model for a specific objective is reused for another task as human beings usually do.
For example, we learn how to use our arm to drink a cup of milk, and then can apply the usage of our arm to drink a glass of wine. If we focus on how to grip a cup or glass, or how to sip, there are slight differences. But we all naturally adopt the core concept to similar tasks.
For computers, it was difficult to achieve learning this kind of generic skill, which is reusable for similar tasks. However, thanks to "transfer learning", we could achieve it.
Now imagine that you train the model for detecting human faces using deep learning. (actually, there are a lot of pre-made models developed by companies or research institute like Google or Stanford) Then you can use the trained model by tuning some parameter and adjust the model for detecting animals since the middle layer of the neural network of the trained model are generic enough to reuse.
Thanks to transfer learning, we might have three benefits:
- higher start
- higher end
- higher learning