This is a... preprint. It's clearly not ready for the prime-time yet. The main interesting finding for me was the utility of variance inflating factors in debu

This is a... preprint. It's clearly not ready for the prime-time yet. The main interesting finding for me was the utility of variance inflating factors in debugging neural networks. But not only the thing lacks a lot of polish (what is their metric for "accuracy"?), but it's still quite "unscholarly", most noticeably lacking: * serious comparisons with classic literature benchmarks (I'm not even saying state-of-the-art) of NNs rather than whatever they were able to find with R. (Who uses R in academic machine learning?) * a better developed argument about how Taylor is good enough. I'm not saying rigorous, but an article about reinforced concrete that just said what they said wouldn't be enough. An example of a workable engineering-type Taylor argument would be something that says that all minima are locally like parabolas and h~=0.01 is a very conservative upper bound for the needed approximation because the material is almost 100 times as strong and we just need 10 times for this calculation. * reproducibility, reproducibility, reproducibility. Don't R people have Jupyter notebooks? Does their Keras package not serialize network structure and weights? Keras for Python does.