May 25, 2020
The question, “What are the problems we should assume can be solved with machine learning?”, or, to be even narrower and more focused on current developments, “What are the problems we should assume a neural network is able to solve?”, is one I haven’t seen addressed much.
There are theoretical frameworks such as PAC learning and AIXI, which at a glance seem to revolve around this, as they pertain to machine learning in general. Unfortunately, if actually applied in practice, these won’t yield any meaningful answers.
However, when someone asks me this question about a specific problem, I can often give a fairly reasonable and confident answer, provided I can take a look at the data.
This piece summarizes the heuristics I use to generate such answers. They are not precise or evidence-based, but I think they might be helpful, and perhaps a good starting point for further discussion on the subject.