It would be nice if there existed a set of standard problems, a set of benchmarks for each of them, and an overview of methods to approach these problems and corresponding benchmarks. Then for each problem, also a set of implementations.
Right now, the field of neural networks seems like a maze. It is too easy to get lost, or to settle on the wrong, suboptimal solution.
This is the point of the benchmarks in literature such as MNIST, CIFAR-10, ImageNet, SVHN, and so on. You can see a pretty comprehensive list here [1] that also shows papers and their reported performance.
There are a lot of implementations for many of these models out there to be found with some google-fu, and usually porting a network from one library or framework to another is not too bad. The main thing is that many modern neural networks are on the hairy edge of research, so having some nice, easy to use code just laying around is pretty unlikely unless the researcher who published the model prioritized making things clean and readable.
The good news is that as long as a "suboptimal solution" is in place in your pipeline, you can always improve later. The hard part is really setting up the pipeline in the first place, IMO.
Since the state of the art is always moving (day to day at times!), and many reported SOTA results are not 100% trustworthy, it is much better to setup a pipeline and test different solutions yourself. One working solution on production data is worth 1000 papers with "optimal" results.
Right now, the field of neural networks seems like a maze. It is too easy to get lost, or to settle on the wrong, suboptimal solution.