|
For Multi-task Learning, the different tasks could learn from each other. While for Transfer Learning, the algorithm can learn the unknown data by the known data of different form. So what is the difference between them? |
|
Transfer learning is used to refer to first learning a predictor for one problem and then transferring it to another problem, and evaluating it on how well it does on this latter problem. Multitask learning is learning the predictors for many problems jointly. See for example this recent (2010) survey on transfer learning, section 2.1, for a definition. Many thanks Alexandre!
(Aug 04 '13 at 22:01)
Ethan Gao
|
|
And I have read a Tutorial from SDM'12, which gives a chart to describe the relationship between them. It is: Multi-class Learning ∈ Multi-label Learning ∈ Multi-task Learning ∈ Transfer Learning |
