Comment by jobigoud
8 years ago
Hmm, current ML has only been worked on for a few years, we have barely scratched the surface. No signs of slowing down afaict. It has wiped the state of the art against algorithms we took decades to think up. What happens after decades of that?
It feels more like there is too many new directions to explore and we don't have enough time/creativity/insights/ideas to fully take advantage of it.
1957 - First ever man-made object launched into space.
1957 - First ever animal launched into space.
1960 - First ever animal launched into space that survives the trip.
1961 - First man goes into space.
1965 - First man goes on a space walk.
1966 - First spacecraft lands on the moon.
1967 - A probe is landed on Venus (!) which collects and sends back video and audio recordings among other data.
1968 - First spacecraft to orbit the moon and then return to Earth.
1969 - We put men on the moon who then safely return.
We've done all of this in just 12 years! We literally went from nothing to men on the moon in 12 years! We've barely scratched the surface of what's possible and there's no sign of things slowing down at all! Can you imagine where we'll be after decades of this. Where will we be in 50 years? I can't even begin to imagine what 2019 will be like.
I don't say this to be snarky, but to point out the issue. Exponential growth doesn't mean continued exponential growth, and it's always very difficult to predict the future. While thoughts such as the above were common place, there would have been hardly a person to predict that the room-sized government "computing machines" would one day be a few inches large, millions of times more powerful, and cost as little as a week or so of minimum wage work. In some cases, those exponents did keep churning for a while.
On AI, I used to be a futurist in terms of it - until I got to work with it. Now I'd be extremely surprised if we have have fully capable self driving vehicle within the next decade that's not constrained to white-listed routes. And that is an extremely softball problem since it fits perfectly into the small domain of problems that current ML/AI systems are really great at making progress on. The thing is getting to a testable model is absolutely trivial. Going that remaining 10% from model to product is many orders of magnitude more challenging.
Current ML algorithms have been worked on for decades earlier. They are gaining momentum again due to big data and computing power.
To me "current ML" in the parent comment meant deep learning. Open a computer vision paper from this month and the prior art section almost only contains references from between 2014 to 2018.
Yes the NN layer architecture might be based on ideas from the previous era, but the way the algo actually solve the problem is completely different.
And that's just because it's the only way we can do it right now. When we can apply deep learning to itself it, to select better architectures and hyperparameters, will find strategies we didn't think about or didn't consider worth trying.
ML was covered in my 1972 AI course at MIT.
And it probably didn't mention neural networks, let alone deep neural networks, which is what the state of the art is using in so many tasks.
I'm not saying it's the end all be all, just that this is what "current ML" is generally meaning, and that this particular approach has only been explored for a few years.