[jampekka]

I would recommend understanding the ML algos at least on a conceptual level. And to use them "raw". It's nowadays quite straightforward with e.g. the transformers-library.

The maths in ANN ML aren't that hard, and you don't need to understand them very deeply even to come up with new models. A lot of the new model development is just stacking pre-built layers with torch.nn.Module.

The difficulty comes from getting the beasts to actually work. But it's largely trial and error for everybody.