A common challenge faced in practical supervised learning, such as medical
image processing and robotic interactions, is that there are plenty of tasks
but each task cannot afford to collect enough labeled examples to be learned in
isolation. However, by exploiting the similarities acr