UPDATE 1: I was trying to be factual in the description of the code without getting too technical and without giving away too many details. What I was trying to say may have been open to misinterpretation as a result. I'm a practitioner (coder) with many years of experience. Writing research papers is just not something that I do. I.e., no interest and no time.
The design goal of the code was to provide good results quickly via a very user friendly API. The robustness and concept drift features were added to make the code easier to use, if that makes sense. Researchers tend to care about 0.0001% better accuracy at the cost of days of runtime on a GPU. I did not set out to compete with them in any way.
I do not work in a lab. I work in the real world. I wrote the code for a specific commercial application. But I can see uses in many other settings. I have not come across any open source libraries that are even close to what I wrote. They tend to be toolkits for machine learning experts.
Several have suggested testing the code against MNIST. Looking into it. I already know the results would be good but not great. I know I cannot compete with an expert using an existing toolkit. But I'll bet my code could easily beat a non-expert using another toolkit. Hope that clears things up.
