Notes on Transfer Learning

Refer from:  http://www.robots.ox.ac.uk/~vgg/publications/2014/Aytar14a/aytar14a.pdf

One vital factor that affect the performance of Transfer Learning is that the similarity of the source data set and target data set.

The key elements of the transfer process: (a) source problems with solutions, (b) a target problem, and ( established transfer connections, (c) established transfer connections. The solution is a computational model which can distinguish the positive instances from the negative ones, in other words, a model can assign a label to a test instance.

The transfer process can be briefly summarized with four main questions: (a) from where to transfer, (b) what to transfer, (d) how to transfer, and (c) when to transfer. The transfer process starts with (i) a target task to be learnt in a target context, (ii) a set of solutions to the source tasks already learnt in the source contexts, and (iii) the transfer (similarity) connections which are decided based on the similarity or resemblance between the target and the source problems.