[sirfz]

I transitioned to what's now referred to as ML engineering 11 years ago, as a programmer I started working on scripts doing all sorts of ETL (in Python - was my intro to the language) and handled datasets saving and loading for training. I managed models serving (we were using Theano at the time) in prod via REST APIs. Also worked (and still do) on writing model architectures with DL experts and I can say I still don't have solid understanding of the maths but I sure know how to manipulate matrices and write ML layers/models when working with an expert.

There are other aspects to the job (like dispatching experiments) but the point is that I was able to bring value in all of this as a programmer without requiring any new skills apart from the natural learning experience that one has to go through in any discipline. I think you can surely transition as long as your job requires it.

[LorenzoBloedow]

I'm looking into learning more about ML and how I can create my own models. I enjoy math, but when I read research papers the formulas sound like gibberish. I also like the idea of Kaggle competitions, do you think I need to have research-level math understanding of ML models to do well on Kaggle?

[sirfz]

Yes, I believe you do need to have a good math understanding to understand why a model is behaving the way it is and develop intuition into where the problem might be but that comes with experience. If you're looking into ML research then absolutely you need to learn the math but if you intend to support ML as a programmer then there's plenty of space for you to do so

[LorenzoBloedow]

Thanks for the response! I'm torn between AI research and just implementing models with my own data, I do enjoy math but I can't seem to understand advanced math, that might change after I get an engineering degree though, or not ¯\_(ツ)_/¯

But for the time being, I'm just wondering if people who win Kaggle competitions implement their own algorithm, or do they just read a lot and try techniques that are already out there?

[sirfz]

Good luck in your studies and plans!

I can't answer you about Kaggle competitions as I never was involved in any but I imagine the novelty would be in the way people encode the datasets and maybe the cost functions they implement but the models themselves are probably based on already established archs