4 ms·
This is an essentially meaningless finding. Neural networks had the dark days they did in the late 90's/early 2000's because computational power didn't allow fo
by rcar 7y ago
This is an essentially meaningless finding. Neural networks had the dark days they did in the late 90's/early 2000's because computational power didn't allow for the depth they needed to be applicable to most problems. Once GPUs caught up to the method, deep neural nets have thrived on some problems.
Those problems are precisely the ones that this paper doesn't evaluate (image recognition, text processing, etc.). In most other domains, tree-based models crush neural nets and statistical regression methods, and so this whole thing just ends up being an academic exercise.
- thanatropism 7y agoIt's not a finding. They never compare their results to literature benchmarks, just to the networks they've managed to estimate in R. They also say that charitably they've refrained to reporting the "worst results", but fail to present a summary of quality of results found over the many networks they ran.
- rcar 7y agoI agree, though they do make the pretty strong statement of NNAEPR suggests that one may abandon using NNs altogether, and simply use PR instead. in the blog post, which sounds to me like they're going for a finding.