Ask HN: How to interpret mathematics in terms of deep learning models?

Ok. First, let me introduce myself.

I am a master's student in Computer Science, in US. I had a lecture on speech recognition today in my machine learning class. It was a good lecture and mostly focused on first principles and basic concepts related to learning speech recognition. Then comes the devil, neural networks jargon, and architecture(LSTM, RNN, Transformers). Till now, all these things make sense to me.

I came back home and gave a thought to everything I learned today. My question is why did we study first principles(like phonetics, n-grams, and all) when deep learning models and all are able to learn patterns themselves. In technical terms, my question is how do we interpret problems, mathematical modeling, and intuition derived from first principles and models in terms of deep learning models and architecture? Because mathematical intuition behind deep learning models looks completely different to me. So, is it just the game of input-output-losses? If possible, if anyone can give me a very classic example of how did you conclude what the deep learning model would look like?

I would ask this question my professor next week. But HN is kind of my professor, so asking here. :P

Thanks!!!