Learning sparse neural networks through L₀ regularization