[m_mueller]

I’ve recently been at a master students’ thesis defense where a method was discussed that is supposed to speed up algorithms by deciding whether to run them on GPU or CPU. The method is to look at the program code with deep learning algorithms.

Fair enough. There’s apparently a set of benchmarks that other papers used and were reused here to compare against. Yet the student was unable to answer my question as to how bias in this dataset was treated or even what percentage of the algorithms was faster on GPU to begin with, they only showed how much better their deep learning fit the data. As in, if 99% is faster on GPU I could as well just provide that as a static answer and be better than any machine learning.

I frankly find it tragic that this kind of stuff is not ironed out in academia. This was was in a very highly ranked university.