I would also mention that the method used to design the new LSTM units is much less interesting than another recent paper on designing LSTM units, using reinfor

I would also mention that the method used to design the new LSTM units is much less interesting than another recent paper on designing LSTM units, using reinforcement learning NNs: Zoph & Le 2016 http://openreview.net/pdf?id=r1Ue8Hcxg http://openreview.net/pdf?id=r1Ue8Hcxg "Neural architecture search with reinforcement learning" (in addition, tested on a more relevant sequence task and performs better, not just similarly).