NeurIPS 2019
Sun Dec 8th through Sat the 14th, 2019 at Vancouver Convention Center
Paper ID:2760
Title:Deep imitation learning for molecular inverse problems

Reviewer 1

The model doesn't seem to generalize very well, especially the AUC for unseen structures is low.

Reviewer 2

The paper is clearly written and motivates the interesting application of finding molecular structures given a spectrum well. The structure of the ms could be improved, since there are some distracting jumps between method, experiments and related work. In particular, the evaluation (Sec 4) could be described in more detail and can be confusing at the first reading. For example, the threshold was only mentioned once before and it could be stated again, that it applies to the spectrum, not the geometry. Here, the paper could also benefit from giving an overview of the training and evaluation procedure, e.g. in a flow chart. The components of the proposed approach - an autoregressive graph model and a predictor of molecular properties (which is unfortunately not described in detail to keep double-blindness) - are known from the literature, but are combined in a new way to solve the structured inverse problem. The presented approach could serve as a blueprint for other molecular inverse design problems, in particular concerning the efficient use of simulated and experimental data.

Reviewer 3

The fast forward model to generate the simulated spectra is not clear, which makes hard to understand the general picture and the theoretical results. This should be described in a clear and extended manner in the supplementary material. Ideally the code should be provided. AUC not defined. In Related work, a summary of references is given but neither substantial analysis nor hard numbers to compare are provided. Line 108-109: Elaborate in the meaning of this sentence to make it clearer, e.g. what is Beta? Missing labels in Fig. 5. Line 203: Forgot to remove note? Did the authors verify that 90.6% was the correct value? Line 222-223: Inverse problems are, in general, a difficult topic and therefore in any application systematic analysis are mandatory. It is not clear how the stability analysis performed really guaranties the robustness of the method. I would suggest a more extended and clear study should be performed.