In this contribution we address the problem of selective transfer of knowledge in multitask learning for classification. We consider selective transfer an interesting framework since when tasks are not truly closely related traditional multitask approaches become suboptimal. We study two multitask learning frameworks where we develop the aforementioned selective transfer paradigm. The two scenarios correspond to two ways of constructing the mapping from examples to output space. In the first naive scenario, the hypothesis is a direct mapping from the input space X onto the output space Y. In the second one, the overall hypothesis space is constructed in a more sophisticate way through a cascade of mappings into intermediate representation spaces. Examples of learning methods under the first scenario are Support Vector Machines (SVM) with linear kernel and Single Layer Perceptrons (SLP). Learning methods under the second type are Multi Layer Perceptrons (MLP) and non linear SVMs.