Regularization and weight reduction

[strakehyr]

Hi all,
I recently noticed that excluding some features from my models would provide better results, even when using dropout in both cases. It weirded me out a bit because I thought the training process would reduce the impact of useless input to 0, but as I have learnt that is never the case. To be able to perform this, one needs to apply L1 regularization. Are there any future plans/existing ways to perform this? I don't know if this is a realistic tool to add or far too complex.