Could the reason machine learning is stuck in a rut be because

>>9126015
no optimization is limited by the amount of information given and ability to fit the information. representing all information to be processed numerically such that it has some complex accurate memory properties is difficult.

no one is teaching in school the next wave of ml right now.

the problem:
https://en.wikipedia.org/wiki/Chinese_room

>>9126018
I think the answer is to symbolize the functioning of the program too, higher order meta cognition. Right?

Pretty much.

The conferences are being flooded with by papers made by teams of asian grad students who just fiddle with conv net architectures until they overfit the dataset even more than the last guy.

Really though, I think the problem is that "Data Science" people all moved over into "Machine Learning" once it became the buzz word. And data science guys are purely concerned with pushing benchmarks, so that kind of spawned this attitude.

Machine Learning used to feel very scientific and mathy, but now its becoming more like engineering.

>>9126281
I agree with this need for higher order cognition.

How ever I think its important not to fall back into the symbolic AI trap of yesteryear. Neural representations can act as fuzzy symbols, and you can still do something very much like symbolic reasoning in a fully neural systems, so I think thats the way of the future.

>>9126018
>https://en.wikipedia.org/wiki/Chinese_room
wikipedia really has dropped in quality

You have to read three paragraphs into the second section before it even explains the theory/thought experiment.

>>9126308
I agree with that. I still see neural interactions fundamentally as sign vehicles, not symbols tho.

How do we make ML great again?

>>9126825
More layers

it's stuck in a rut because the field is full of code monkeys publishing minor variations on network structures that they happened on that fit some small minor specialized problem