Popular convex approaches for sparse estimation such as Lasso and Multiple Kernel Learning (MKL) can be derived in a Bayesian setting, starting from a particular stochastic model. In problems where groups of variables have to be estimated, we show that the same probabilistic model, under a suitable marginalization, leads to a different non-convex estimator where hyperparameters are optimized. Theoretical arguments, independent of the correctness of the priors entering the sparse model, are included to clarify the advantages of our non-convex technique in comparison with MKL and the group version of Lasso under assumption of orthogonal regressors.
展开▼
