Arel’s neuroscience inspired AGI
by Shane Legg
The Singularity Summit ‘09 videos are now up and I’ve been asked about the relationship between Itamar Arel’s talk and the neuroscience part of my Halloween talk (1 minute into Part 9, through Part 10, Part 11 and Part 12).
The short answer is: yes, our perspectives are indeed very similar. Essentially, brain-like deep belief networks + brain-like reinforcement learning + powerful computers = AGI quite soon. This similarity isn’t all that surprising: I know a number of people who are thinking along these lines. I actually met Itamar briefly before the conference and mentioned that I did reinforcement learning, but it was only after his talk that I realised how close our perspectives are.
Given the obvious similarities, what are our differences? The main difference seems to be our perspectives on the maturity of deep belief network algorithms. I think these algorithms are quite impressive, but I think that it will take another decade of research before we are ready to even attempt human level AGI, and then something like another half a decade before we will have a realistic chance of getting there. He thinks the technology is more mature and we’ll get there in about half the time.
Comments
I’m not an AI researcher but just a computer engineer with an interest in the topic who is trying to educate himself.
Well, Itamar Arel where he said that with sufficient funding (about ten million if I remember correctly) he thinks we have good chance to build an AGI system in 3-5 years.
I’m sure this guy knows his stuff and is very smart but I just can’t believe we are really so close.
Just saw a Jeff Hawkins conference video ( by the way, what do you think of his approach?) and he says there is a prize in computer vision.
One million dollar prize for building a system which can correctly solve a classification problem.
The system must be able to discern pictures of dogs and pictures of cats. Even a 3-year old human can easily do that.
If that’s our current state of the art in computer vision I just can’t believe we can progress so much in 3 years to be able to build an AGI.
Unless Itamar already knows many things that the narrow-AI community ignores.
If I remember correctly, his claim in his talk was that he could do it within 10 years. Like you, most people don’t consider that to be realistic.
My claim, which most people also think is over optimistic, is that in 10 years time we should know enough to be able to start a project like Arel’s and have a significant probability of succeeding within a decade.
Regarding Hawkins. There isn’t really anything new in his book, it’s all standard stuff that is well known in neuroscience. The good thing about his book is that it has made more people in artificial intelligence aware and interested in neuroscience, in particular, there is a stronger feeling that AI people might be able to learn useful things from neuroscience. Since the book, Hawkins has put all his energy, or so it seems, in a Bayesian network model rather than following a more neuroscience direction. This type of network is pretty standard stuff and plenty of others are and have been working on it for years. There is nothing much new, as far as I can tell, in what he’s doing here. I think his work would be been more interesting if he’d tried to stay more in line with the direction he was going in his book, essentially trying to come up with an abstract computational model of the cortical column algorithm.
Thanks for your reply.
I was interested in Hawkins approach but of course I don’t know what is new and what we already know about the brain.
Yes, Itamar Arel said 10 years in the talk, I saw the 3-5 year timeline on an interview I read on nextbigfuture.com.
Like you (and Hawkins) I always thought ( intuitive and qualitative ideas from someone who has not a deep knowledge of the domain like me are almost always naive at best and stupid at worst, but anyway… ) we should get inspiration from the brain.
We want to build an AGI machine and we don’t know how to do it.
The brain is the only working AGI machine that exist so it makes sense to start copying it.
I also think (and I’m not the only one since I read this opinion somewhere) the brain is full of not useful (from an AGI perspective) things like controllers of heart and respiration and also full of “legacy code” accumulated through thousands of years of evolution which in many ways is an accidental process.
So we can do better, but only after the Eureka moment of fully understanding brain working.
I just looked that interview up… thanks. Yeah, I was meaning human level AGI, not baby level AGI, just to be clear.
The brain is a really complex thing and there is a lot we don’t understand about its functioning. Even some really basic stuff about how synapses change over time is very complex and not well understood. This is the first thing to know. Nevertheless, there is a lot that we do know about the brain, and some of what we do know points in directs for how we might go about designing an AGI.
Unfortunately, we don’t understand perhaps the brain’s best trick: how it builds the cortical abstraction hierarchy. This is what Hawkins talks about through most of his book. In my opinion, if we can figure this out, either by understanding it in biology or coming up with something from machine learning that fills the same role, then building and AGI won’t be all that hard. Will we figure this out in the next 10 or 20 years? Maybe. Maybe not. My guess is that we will, but that’s just a guess.
Regarding legacy code. Yes, some of the brain is wired up to do all sorts of things like breathing and blood pressure… stuff that an AGI wouldn’t need. In some cases things are mixed up, however, so it’s not so clear cut. For example, the hypothalamus is involved with lots of things like blood pressure and breathing, but it’s also involved with dopamine and memory.
In some cases things are mixed up
Oh, Yeah?
Living “mechanisms” are mixed up?
What a surprise!
If only the Wright brothers had spent more time studying the structure of feathers, the fine arrangement of bones and muscles in a bird breast may be we could be closer to Heavier Than Air Flight…
I don’t mind criticism, but you if don’t dial back the sarcasm I’ll start deleting your comments.
I’ll make my stance clearer:
- Not much hope for AGI thru sequence prediction, AIXI, solomonoff induction or any other math gimmicks because this is too simplistic, before you come to the point where math matters you have to figure out the objects your maths are about.
Sort of an “ontological” problem, but as you may (not?) know I am not a Platonist either.
The stage where objects and concepts are somehow “created” is entirely missing from the AIXI approach which only cares about events about them (a concept!) thru time (a concept!) and space (a concept!), etc…
Thru which hole does the electron pass in the double-slit experiment?
This is a meaningless question, one need a complete rehaul of the conceptual framework to work on such problems (Schrodinger & als did it!).
- I find the “hope” about reverse engineering of the brain even more misdirected.
As you certainly know the brain is a HUUUUGE complex mess, nothing like a blueprint, before anyone can have Ha Ha experience about the high level functionality of this or that part of the brain they will drown in a morass of biological kludges akin to the ones I linked to.
Our personal brain power would certainly be put to better use on other paths.
This was the essence of my sarcasm.
From my perspective neuroscience is providing a growing list of insights into how the brain is designed — insights that I think are very useful for AGI design.
You mention the importance of concepts. From my perspective concepts are essentially just high level abstractions in the cortical hierarchy. I think that producing these increasing levels of abstraction is the unsolved key to getting AGI to work. Unfortunately, neuroscience isn’t currently making much progress in explaining how cortex manages to do this.
I also don’t “think” (maybe I should say “feel”) it is possible to achieve real AGI through AIXI (Shane thanks to your Halloween lecture I think I finally got the point of it. After that everything I read about AIXI and algorithmic learning theory made sense ).
Maybe some useful systems for proving theorems or learning to play many different board games or some others not so narrow but not totally general purpose systems.
But defining intelligence is useful to set some kind of standard and to tell other people “look, now we have a clear goal”.
I also think the crucial problem is related to knowledge representation or frame reference, basically how to create and update a model of an unpredictable environment.
How to do that in a way that’s computationally cheap.
How to know what is relevant and what’s not.
How to acquire commonsense knowledge.
The fact is that’s a problem we don’t know how to solve.
So it makes sense trying to see how the brain solves it.
Not whole brain reverse engineering, just learning some of the best brain tricks has you already said.
Shane I see you have worked with people who are in the field of AGI like Ben Goertzel and Peter Voss. From what I understand they are already coding and building systems.
Can I ask how close are they to the goal?
In my opinion: they aren’t close.