It was realized soon after the introduction of Multiple Kernel Learning that ℓ1 - regularization and its groupwise extension are related to MKL by dualization. MKL can therefore be viewed as providing an extension of sparsity to function spaces. However, this extension is not limited ℓ1 - norm. For example, extensions to ℓp norms leads to several forms of non-sparse MKL. In this talk, we will discuss how structured sparsity can generically be introduced in the MKL framework, how existing algorithmic approaches extend to this case and how this leads naturally to the notion of structured functional spaces.