Like they say in the article, there's general AI (think sci-fi computers with minds) and specialized AI (think good old fashioned statistical models, but applied to more things and super effective).
Specialized AI (and I hate calling it AI) is coming along really quickly. We're getting better at it in existing fields and learning to apply it to new fields. More than anything, we just have so much data on everything now, and computers are pretty powerful now, so even old school models are finding tons of new applications.
Generalized AI is a different story. We are a few really major breakthroughs away. We aren't even 100% sure if they are possible, much less understanding how to do them. These aren't the normal slowly-chip-away-at-it breakthroughs, these are things we have no clue about. With something like that, who can really say how far we are? It could be 5 years, it could be never.
We don't even have good generalized robot manipulation. Think of that as monkey-level AI. After half a century of hard work, robot manipulation in unstructured situations still sucks. Go watch DARPA Humanoid Challenge videos or old Willow Robotics towel-folding to see how badly it sucks. Factory robotics works because the work situation is highly structured.
On the other hand, once someone cracks that, a huge number of low-end jobs will be automated.
Also mentioned was that specialized AI is creating social problems we need to solve. There are issues with fairness—specialized algorithms all too easily replicate bad/unfair patterns in the data (e.g. they replicate/amplify unfairness in the world), and that harms folks; sometimes they might not even be aware of it.
That's exactly what is about to happen. Deep learning has the potential to do robot control. Currently researchers are beating tons of video games using reinforcement learning with deep networks. Applying the same methods to robots shouldn't be too hard. And we've also come a long way with machine vision as well over the past 5 years.
I'm as bullish on AGI as anyone in the medium term, but deep learning is not even playing the same game as AGI, let alone in the same ballpark or having the potential to achieve it.
Deep learning is still mere perception. It doesn't handle memory or processing, it just transforms input into output, typically trained by Big Data, way bigger than necessary statistically speaking, given the world we live in.
AGI requires super aggressive unsupervised learning in recurrent networks, likely with specialized subsystems for episodic and procedural memory, as well as systems that condense knowledge down to layers of the network that are closer to the inputs. At a minimum. And nobody is really working on any of that yet (or at least succeeding) because it's really damn hard.
That's why everyone in "AI" is rebranding as a deep learning expert, even though deep learning is really just 1980s algos on 2016 hardware - you gotta sex up feed forward backprop or you don't get paid.
Edit: to be fair, robot control is much simpler than AGI, and might be mostly solved with deep learning somewhat soon, I forgot the context of your post.
Sure, and I probably shouldn't have glossed over that. That sort of research is definitely progress, though it's not paradigm shifting in any way. I do think that we are getting past perception slowly but surely, I just don't think we're there yet.
What really doesn't exist is any meaningful stab at unsupervised (or self-supervised) training on completely unstructured inputs or any sort of knowledge condensation/compression, at least for time dependent problems. These are of paramount importance to the way we think, and to what we can do.
There's a lot of trivially low hanging fruit, too - I still have yet to see even a grad school thesis that starts with an N+M node recurrent network and trains an N node subnetwork to match the outputs based on fuzzed ins, and then backs that out into an unsupervised learning rule that's applicable to multiple problems. Or better, a layered network that is recurrent but striated, that tries to push weights towards the lower layers while reproducing the same outputs (hell, even with a FF network this would be an interesting problem to solve if it was unsupervised). These are straightforward problems that would open up new avenues of research if good methods were found, but are mostly unexplored right now.
I could be wrong, if I had real confidence that we were close I'd be working on this stuff, but I'm collecting a paycheck doing web dev instead...
Sequence predicting RNNs are basically unsupervised, in that they can learn from lots raw of unlabelled data. And they learn useful internal representations which can be adapted for other tasks. There is lots of old work on unsupervised learning rules for RNNs, including recurrent autoencoders and history compression.
> we introduce a form of memory-augmented neural network called a differentiable neural computer, and show that it can learn to use its memory to answer questions about complex, structured data
So it seems that deep neural nets can have memory mechanisms and be trained to solve symbolic operations.
I'm not talking about AGI at all! Just robot control. It's a difficult problem sure, but not that difficult. There has been massive progress on it, and related problems. I have no doubt we will have 'solved' it in at a decade.
The systems which learn video games work only for games where the state of the game is entirely visible, and the desired player action can be decided based only on the current state. PacMan, yes. Doom, not that way.
That's only because they didn't use recurrent neural networks which save information over time. RNNs make it possible to play games with hidden state. Deepmind is currently working on that with starcraft, which is vastly more complicated than pacman. They also have some work on 3d games like doom.
A few weeks ago there was a paper posted on "Synthetic Gradients" which should make it much more practical to train RNNs for games. Before it required saving every single computation the computer makes to memory, which uses a huge amount of memory and computation. Using synthetic gradients they need only store a few steps in the past. And it can learn online.
The ability for AI to approach these problems has only been possible in the last 2-3 years. The tech was really not there in 2008, and it's still very rough and cutting edge in 2016. But we are at least seeing the first glimpses that it's definitely possible. If AI can play starcraft, then surely it can control a simple robot. And anyway see my other comment.
Honest question: Why do you think that generalized and specialized AI are distinct things? Is it not possible that general AI is a specialized AI applied over the field of specialized AI generation?
What people mean by generalized and special AI is not consistent in the field, but everyone agrees that the current brand of AI driven by statistical learning techniques and large scale neural networks is far from explaining how even the simplest of nervous systems work. The key obstacle is adaptation. Several people believe that they've more or less solved the recognition problem. However, adaptation is a totally different thing. There are no tools in the current AI toolkit that we can use to make a robot that can go out unsupervised in the real world, do something useful and come back safely. Whereas even the Nematode C. elegans with only 302 neurons is remarkably flexible, it can forage for food, remember cues that prediction food, manage food resources, get away from danger or noxious stimuli etc. This allows it to survive quite well in a world that is constantly changing in unpredictable ways. This is the kind of intelligence that proponents of so called general AI want, and I agree we have a couple of major breakthroughs away.
And we have a complete wiring diagram for C. elegans, and no clue how it does any of the things you talked about. So, yeah, general AI is really far off.
To be honest, the wiring diagram is a bit of a distraction from the really big questions. It has its uses for sure, and is really essential in many situations but overall it gives this illusion that we understand something important about the system, where in reality we don't. Understanding a biological system from its wiring diagram is something like understanding a city by studying its road map.
I have no idea what I'm talking about. But why couldn't we build some sort of bio-computer hybrid system around a simple form of life, like "C. elegans" but augmented with traditional CPUs?
> Is it not possible that general AI is a specialized AI applied over the field of specialized AI generation?
AI problems can be characterised as those where there's no clear path to a solution (otherwise we just call it "programming"); tackling them necessarily involves trial-and-error, backtracking, etc.
Since there are far too many possibilities to enumerate, solving such problems requires reasoning about the domain, e.g. finding representations which are smooth enough to allow gradient descent (or even exact derivatives); finding general patterns which will apply to unseen data; finding rules which facilitate long chains of deduction; etc.
The difficulty is that there's usually a tradeoff between the capability/expressiveness of a system, and how much it can be reasoned about. If we choose a domain powerful enough to represent "the field of specialised AI generation", for example turing machines or neural networks, methods like deduction, pattern-finding, gradient following, etc. get less and less applicable and we end up relying more on brute-force.
To me, this is where the AI breakthroughs are lurking. For example, discovering a representation for arbitrary programs which allows a meaningful form of gradient descent to be used, without degenerating into million-dimensional white noise; or to take deductive knowledge regarding one program and cheaply "patch" it to apply to another; and so on.
My two cents: they are separate because there is no current algorithm that can take us from modeling (whether classical statistics or neural net) to intelligence. Applying our current specialized techniques to AI generation has not gotten us there. That is because the techniques are mostly model tweaking techniques. The models are generated and trained for each problem domain. A combined solution may be developed soon, but I doubt it.
There was a great article recently on HN that highlights the current problems:
Just because we may acquire the processing power estimated to be used in the brain (in operations per second) doesn't mean we know how to write the software to accomplish the task. It is very clear current algorithms won't cut it.
Also, I think we are a few orders of magnitude off on raw processing requirements because I think it is a bandwidth issue as much as an operations per second issue.
TL;DR - you could throw as much processing power and data as you want at any current deep NN or their derivatives and you wouldn't get general intelligence.
That said I don't think the winter will be as bad as before because, like OP says, specialized AI is useful.
Specialized AI is all about X,Y pairs. Given X, predict Y. There are other problems it's good at too, like given X, choose a Y to optimize Z, but at it's core it's largely the same. On the fringes, you have stuff about exploration, which is AWESOME, but still pretty niche. At least 99% of the "AI" you hear about is of the X,Y variety. More to your point, if we can make generalized AI from "given X, predict Y," then nobody's figured out how to do it, and nobody has super promising research tracks to get it.
I think a lot of the early AI research (not my specialty) had the idea that if we made a bunch of systems that were good at their own piece of the puzzle, then we could just tack them together and get real intelligence. It just didn't turn out that way. Something I'm more familiar with is graphical models, and while they in principal could do amazing things when you stick little expert components together, we've proved the complexity grows pretty badly in the most general cases that would have been really amazing. I'd bet similar things happened in other "let's put a bunch of specialized systems together" tracks. Maybe we can do it, but not the naive way that would have been great.
Then you can get interesting and philosophical about it, where you might even say that emulating intelligence and intelligence are different. Like the chinese room thing, or even a character in a story vs a physical person. I'd rather not weigh in on that right now, but there are good interesting arguments both ways.
I guess I should have been more specific. I meant sorta convincingly emulating intelligence versus fully meeting some other definition. Is a turing test enough?
This is not a problem because there is no such thing as generalized AI. There's just specialized AI for lots of things added together. Keep getting better at things and after a while you are good at lots of things, so appear to be general to an observer (until they see you outside your domain of expertise).
This theory compactly explains why no one knows how to do general AI.
I don't believe there is a special sauce waiting to be discovered.
That's not what General AI is though (at least how I define it), real AI would have to be able to invent new things, not just be good at existing things.
Many humans go their whole lives without doing that, so I'm aware it's a high bar. But it's a bar that some humans do pass, and if AI is to be more than just a helpful gimmick, it'll have to do that as well, since I'd like to believe all humans have that potential, even if not always realized.
(Obviously a helpful gimmick still does have value.)
But brain still is "just" a neural network, granted, with immensely more complex neurons and extraneuronal mechanisms that might as well be absolutely crucial for learning. But the difference between a highly intelligent and blank-stare person can be as small as switching off a small part of the brain or tweaking a neurotransmitter. Which shows me that general intelligence is something extremely sensitive as oposed to something like vision processing, which takes much more drastic changes to disturb.
Looking at the brain, it often does look like a bunch of interconnected specializes neural networks.
Specialized AI (and I hate calling it AI) is coming along really quickly. We're getting better at it in existing fields and learning to apply it to new fields. More than anything, we just have so much data on everything now, and computers are pretty powerful now, so even old school models are finding tons of new applications.
Generalized AI is a different story. We are a few really major breakthroughs away. We aren't even 100% sure if they are possible, much less understanding how to do them. These aren't the normal slowly-chip-away-at-it breakthroughs, these are things we have no clue about. With something like that, who can really say how far we are? It could be 5 years, it could be never.