I started learning ML especially deep learning couple of years ago. Worked in grad school research lab. Gained experience in the field, published a paper. But I'm not able to appreciate this field. Learning statistical machine learning is more fun than deep learning where it is just the matter of trail and error and Coming up with the "novel" architecture. I don't see mathematical rigor there. On the other hand I feel the engineering aspect of it more challenging and interesting which involves distributed training But I see people going crazy when the word deep learning is being used. Am I missing something here?
It’s a bit of a joke . But a joke that works
Its funny when people use the term AGI will be reality with deep learning on something which no one know why it works and how it works 😒
I think we’ll have AGI before we know how it works. Intelligence might be too complex to reason about in simple mathematical terms.
You’re overfitting bro
You nailed it.
Blind allows "nailed" now? Yay!
Think of it as a soft science, like psychology.
I had this recent realization that I am not good with AI/ML field where you try to figure out data .. I rather be what I do now... backend software engineer .. reason for any decision is much more clear
I understand the sentiment behind your post OP, but industry only cares about results, not about the multitude of different, and perhaps more interesting, approaches. Deep learning in many cases brings results.
It's an inexact science, and toolkits and frameworks allow aggressive, untrained people wing their way into scientific domains in large numbers, overwhelming those who understand the need for rigorous approaches. If science were done like software, we'd still be in the 18th century with no standards for electricity, telecom, transportation, etc.
Right now the industry has tons of pretenders bullshitting their way in ML/AI it's infuriating.
those that talk about deep learning, know nothing about deep learning. statistical ML is the only way to go!
I have done deep learning. I left and joined as SDE. Deep Learning is less logical. You can always blame it on bad data. The model basically memorizes. If Deep Learning was not just that then Tesla would be self driving on all kind of roads.
You mean the difference between something working and not working boiling down to using a sigmoid vs a relu, with no real clue why? Yeah, I feel you.
Exactly. Filter size, number of neurons, number of layers.. I can go on I get it that these are architecture decisions but unfortunately without a foundation 🤷♂️
I know. Still magical when it does work though.