Over the past two years or so, I have noticed a steady rise in the fraction of new questions - especially Python questions - about some machine-learning problem, wherein the OP is trying to improve an AI model that functions without error, but gives low-quality output. Because it is an AI system operating on a possibly vague task, it may be hard to say objectively that the output is wrong per se, but people will report symptoms such as apparent overfitting or underfitting, attempts at gradient descent that don't "converge" well, unexpectedly low or high "accuracy" self-reported by the model, etc. (Disclaimer: I have at best a very rudimentary understanding of these concepts.)
It appears that in most if not all of these cases, the OP is expecting that the problem should be solved by tweaking some input parameters to the model (most of which is some heavyweight library abstraction), rather than by finding and fixing some problem in the actual logic of the code. Note that I am not talking about questions where there appears to be some logic error in the code unrelated to ML, nor about e.g. ValueErrors raised by Tensorflow when the input data array for a layer has an invalid shape.
Are these questions suitable for Stack Overflow? If not, what closure reason is appropriate?
Can these OPs be pointed somewhere else on the Stack Exchange network?
Can this sort of question be genericized in any useful way? Would it ever be viable to close one question of this sort, as a duplicate of another?
