Fast ℓ1-minimization and parallelization for face recognition

While ℓ₁-minimization (ℓ₁-min) has recently been studied extensively in optimization, the high computational cost associated with the traditional algorithms has largely hindered their application to high-dimensional, large-scale problems. This paper discusses accelerated ℓ₁-min techniques using augmented Lagrangian methods and its parallelization leveraging the parallelism available in modern GPU and CPU hardware. The performance of the new algorithms is demonstrated in a robust face recognition application. Through extensive simulation and real-world experiments, we provide useful guidelines about applying fast ℓ₁-min on large-scale data for practitioners.