Gap Safe screening rules for faster complex-valued multi-task group Lasso

Mathurin Massias, Joseph Salmon, Alexandre Gramfort

2017-01-01

Abstract:Linear regression with sparsity-inducing penal-ties is a popular tool for high-dimensional inverse problems such as source localization. The Group Lasso is a particular choice of penality that considers the l2, 1 norm to promote group (or block) sparsity patterns. Since no general closedform solution is available for this problem, iterative solvers are needed. This can lead to very slow convergence especially if the problem is ill-conditionned or if the dimension of the problem is particularly large. Safe screening rules [3](see also [12]) and in particular dynamic ones [1, 2] speed-up the optimization process by progressively discarding regressors identified as irrelevant. In this work, we consider the case where features and observations can be complex-valued, a common case in signal processing when working with timefrequency operators. We derive Gap Safe screening rules in this context and propose a block coordinate descent optimization strategy [11]. In practice, we illustrate significant speed-ups in terms of convergence compared to classical solvers on a neuroscience problem, namely the problem of source localization using magneto and electroencephalography (MEG/EEG).