NeurIPS 2020

Linear Dynamical Systems as a Core Computational Primitive

Meta Review

The paper proposes to use linear dynamical systems as a core primitive for building recurrent networks that are more interpretable and easier to analyze. The paper received mixed reviews (below acceptance, top 15% -> top 50%, and top 50%). On the positive side, the approach is novel and the theory well developed. On the negative side, some parts are not clear and the experiments need improvement. Overall, the paper is a stepping stone toward developing recurrent architectures that are better connected with classical dynamical systems, more interpretable and easier to analyze. Thus, I believe this paper offers an interesting direction for future research in machine learning.