Reviews: Architectural Complexity Measures of Recurrent Neural Networks

The goal of the paper is to provide a theoretical and quantitative guideline for designing an RNN for a specific problem. This is an important and hard topic that worth investigating. The measurements of d_r (Recurrent depth), d_f (Feedforward depth) and s (Skip coefficient) seem to be reasonable. The authors also conduct comprehensive experiments to compare several RNN variants. However, I find the paper hard to understand.