{"url": "https://www.youtube.com/watch?v=IjG_Fx3D0o0", "page_url": "https://www.youtube.com/watch?v=IjG_Fx3D0o0", "title": "Demis Hassabis \"Systems neuroscience and AGI\"", "show_or_channel": "Future of Humanity Institute (Winter Intelligence 2011)", "date": "2011-01-16", "duration_s": 2419, "type": "talk", "speakers": ["Demis Hassabis"], "confidence": "high", "notes": "FHI Winter Intelligence Conference, Oxford: \"Combining Systems Neuroscience and Machine Learning to Build AGI\".", "people": ["demis-hassabis"], "id": "2011-01-16_future-of-humanity-instit-demis-hassabis-systems-neuroscience-and", "audio_duration_s": 2418.0564375, "worker": "jc-w21", "transcribe_s": 109.3, "model": "large-v3", "segments": [{"start": 0.0, "end": 25.84, "text": "I'm going to be speaking about systems neuroscience and how it might be used to help us build AGI. And really my talk splits into sort of two parts. So firstly, I'm going to make the case for why I think that systems neuroscience might have an important role to play. And then I'm going to sort of talk about, you know, AGI, as we know, is an incredibly hard problem."}, {"start": 25.84, "end": 55.36, "text": "And reaching human-level AGI, even by the most optimistic kind of estimates, is probably a 20-year-plus project. So what kind of interim goals might we expect to see on the path towards AGI? And sort of, you know, a related question is how can we know, how can we measure whether we're making progress towards this path? And, you know, what is it we can do to kind of aid that?"}, {"start": 55.36, "end": 84.82, "text": "Before I talk about systems neuroscience specifically as an approach to AGI, I'm going to take a sort of top level view of the spectrum of approaches taken in the past and present actually towards building AI. So at the very top level, the top level kind of distinction I would say is between non-biological approaches and biological approaches. And in fact, the non-biological approach was taken first and really"}, {"start": 84.82, "end": 113.55, "text": "is what is being referred to by other speakers as good old fashioned AI. And really what that means is the kind of symbolic AI approaches of the 70s, 80s, and 90s, which include things like formal logic systems, logic networks, lambda calculus, expert systems. So the whole range of techniques were tried. And they're all kind of mathematically or logically based, kind of grounded. And the general problem with all of these systems and why they didn't"}, {"start": 113.55, "end": 141.89, "text": "They tended to suffer from one or more of these problems. They tended to be brittle, didn't deal very well with ambiguity or uncertainty. They were time consuming to train, or perhaps in light of Moshe's talk, we could talk about needing huge amounts of data. They were pretty poor at generalizing. It was difficult for these symbolic systems to acquire new symbols, to generate new symbols that they didn't already know."}, {"start": 141.89, "end": 171.9, "text": "And actually Ben's already talked about earlier some of these classic problems like the symbol grounding problem where if you have a system that's just describing entities with other symbols, how do you actually refer to things outside of itself, outside of the agent? So really this was taken to its extreme in the Site project which was led by Doug Lennert."}, {"start": 171.9, "end": 179.76, "text": "Sorry, is even, yes. I should say. It's true, true. Ongoing site project where..."}, {"start": 179.76, "end": 202.91, "text": "Really, this was taking it to its limit where I think Lennox proposed that the real problem with these symbolic AI systems was not with the approach, but the fact that there wasn't actually enough information in these databases. So he went about trying to solve this by attempting to put every piece of common sense knowledge and how he defines that into a..."}, {"start": 202.91, "end": 231.09, "text": "One massive database, the site database. And as Moshe points out, it's still ongoing. 25 years later and still no sign of any intelligence. So on the other kind of side are the biologically inspired approaches. And really, this kind of makes some kind of intuitive sense because there are 6.8 billion, or however many humans there are alive now, examples of AGI systems out there, which is the brain."}, {"start": 231.09, "end": 260.99, "text": "And so it kind of seems to make intuitive sense that maybe we should be thinking about using the brain as a blueprint. But even on this side of the divide, this actually covers a very large range of quite different approaches. Now, we can argue about which of these two approaches makes most sense. But actually thinking about this, I think what this really boils down to is really your personal intuition about what the search base of"}, {"start": 260.99, "end": 282.77, "text": "Possible intelligences looks like. So on the one hand, if we think about the search space of possible AGI solutions, we could be in a regime where the search space is quite small and there are a large number of possible solutions. So it's a dense search space. So pictorially, this cartoon."}, {"start": 282.77, "end": 311.95, "text": "Kind of illustrates that. So if the outer circle is a container is the search space, the entire search space and the stars are possible AGI solutions. And then the human brain in this kind of scenario is not that special. And perhaps it isn't worth the time and effort and resources to take and to understand how the brain solves these problems. On the other hand, we could be"}, {"start": 311.95, "end": 343.23, "text": "In regime two, which is a relatively large search space that has relatively sparse number of solutions. So we might be in this kind of regime. And in this case, it would seem to make sense that we should pay attention to the human brain and spend effort into understanding how it solves some of the harder problems rather than just approaching some other way of searching this very large search space. So which regime are we in?"}, {"start": 343.23, "end": 367.07, "text": "I mean, it seems to me, you know, so there's regime one, small and dense, or regime two, large and sparse. And there seems to be two pieces of evidence, I think, that points towards the fact that we're actually in regime two. One is a natural kind of argument, which is that evolution has only produced human-level intelligence once, after, you know, billions of years of trying. And probably more personally, one of the..."}, {"start": 367.07, "end": 397.54, "text": "The big things that the projects in the 80s and 90s showed is that intelligence is a very, very hard problem and was severely underestimated, I guess, in the 60s and 70s. And they've largely, although they've yielded interesting results, and in some cases very interesting applications, they've largely failed to make progress towards the overall goal of AGI. So then, if we kind of, for the moment, buy this argument that a biological approach to AGI"}, {"start": 397.54, "end": 428.54, "text": "Make some sense. Then we can subdivide this further into different categories. So here's a kind of spectrum of biological approaches from on the left-hand side, the most abstract, i.e. least tied to the biology, and on the right, the most biological. And I'm going to position various well-known projects and approaches on this scale. So I think over here on the left-hand side, the most abstract and kind of most loosely based on"}, {"start": 428.54, "end": 449.74, "text": "How the brain does things is what's called the cognitive science architectures. And I could have listed actually dozens of possible examples of projects here. And I've just listed three of the more well-known ones here. There's SOAR, which is Alan Newell's project, and John Lead's. ACTAR, which is from John Anderson. And more recently, Ben Goertzel's OpenCog."}, {"start": 449.74, "end": 478.05, "text": "What these mostly have in common is effectively you can display them or think of them as box and wire diagrams. This is actually a diagram of ACT-R. What you see here is the boxes represent modules very loosely based on brain regions or brain processes that are just wired up together to overall create what's hopefully some kind of intelligence."}, {"start": 478.05, "end": 503.76, "text": "The main reason why I think there are so many actual approaches on this front is that generally speaking, these are based on effectively introspective processes on the parts of the designers of these systems. And that's what I think kind of makes them a little bit unsatisfactory is in general, there aren't deep underlying principles as to why an architecture is being designed in this specific way. And you can see that"}, {"start": 503.76, "end": 534.29, "text": "And then all of these architectures have to sort of retrofit that into their existing kind of wire diagram. So it feels like a bit of an unsatisfactory, unprincipled approach to AGI. And on the other hand, we have the very biological approaches."}, {"start": 534.29, "end": 562.21, "text": "Things like BlueBrain and Moda's Synapse projects. And really we can think of this as close, akin to the whole brain emulation camp. And what this camp proposes to do is really look at copying or mimicking the brain to a very fine grain detail. So here actually I'm showing a beautiful diagram from one of Moda's recent papers showing the wiring of a macaque monkey brain."}, {"start": 562.21, "end": 587.79, "text": "And these are all the main pathways in the macaque monkey brain. Now, the problem I have with this approach is that this is a beautiful diagram and it has a lot of uses in biology, but how much is this wiring diagram really telling us about processes or the underlying functions the brain is supporting? The other issue I have with these kind of whole brain emulation approaches is that they rely on very..."}, {"start": 587.79, "end": 617.04, "text": "Intricate imaging techniques. And I think that to get down to the level of detail we're going to need to do something like whole brain emulation, and what level of detail we go down to is an open question. Is synapse enough? Do we need to go to calcium channels? The atomic level? What level do we stop at? Is actually that our imaging techniques are still decades away from having that kind of sophisticated imaging techniques, especially if we want to keep the original brain or the original copy intact."}, {"start": 618.06, "end": 645.55, "text": "So what I'm suggesting here is actually there's a sweet spot in the middle, which I'm calling the systems neuroscience approach. And where we need a sort of, what we're really interested in is the algorithms the brain's implementing. So not the specific implementation details, but the algorithms and representations. And what this can map onto is David Mars, who is the kind of founding father of computational neuroscience, his three levels of analysis."}, {"start": 645.78, "end": 674.13, "text": "He outlined in a classic paper in the 70s. And what he suggested is that to fully understand any complex biological system, for example, the brain, we need to understand it simultaneously at three different levels. So there's the computational level, the highest level, algorithmic, and implementation level. These are the three levels he called them. Now, the computational level is really the what. So the goals of the system. What is the system trying to do?"}, {"start": 674.13, "end": 703.46, "text": "The algorithmic level is the how, so the representations and the algorithms, how it achieves those goals. And then finally, the implementation details, so the medium, the substrate, the physical realization of the system. Now, if we now compare these levels of analyses that Mara suggested with that spectrum I showed you in the previous slide, then we see they map on quite well. So the whole brain emulation camp would be really focused on the implementation level."}, {"start": 703.46, "end": 726.45, "text": "How does a system physically carry out intelligence? On the other hand, the cognitive scientist architecture approaches are really more at the computational level. And therefore, what we're talking about focusing on here is really this in-between level, the algorithmic level, the how, the representations and the algorithms."}, {"start": 728.37, "end": 757.49, "text": "Now, that would all be very well saying that we should use systems neuroscience as an aid, as a source of inspiration for AGI ideas, but that would have been kind of a moot point until around 10 years ago, or maybe you could say 15 years ago, where really cognitive neuroscience has rapidly taken off. So in the last sort of 15 years, and especially the last 10, there's been a revolution in cognitive neuroscience. I'll just show you a couple of graphs here. I don't have a laser pointer."}, {"start": 757.49, "end": 769.82, "text": "This top graph here is actually a graph of number of citations in tens of thousands of papers against year. And you can see that around 1995, there's virtually none."}, {"start": 769.82, "end": 796.8, "text": "And here we go, we're up to sort of like 100,000 now in 2008. And also, it's not only exponential in total, the newer technologies which are represented by the blue, red, and green. So green represents fMRI technology, which is the latest one. Red is PET, and blue is MEG, which is the oldest technology, EEG. You can see that the proportion of the newer technology is growing at the same time as the total."}, {"start": 796.8, "end": 826.59, "text": "There's been a massive revolution in cognitive neuroscience and there are really new experimental techniques coming online every year. So recently we've had optogenetics where you can actually have cells firing when you shine a laser on them, TMS where you can actually disable, non-invasively disable a part of the brain, a human brain, two-photon microscopy, all sorts of things. And on top of that, compounding these kind of advantages and new techniques."}, {"start": 826.59, "end": 855.57, "text": "have been new analysis tools taken from the machine learning world. So for example, multivariate pattern analysis and support vector machines, all kinds of latest machine learning technology. And all this wrapped in together has resulted in a kind of exponential growth in understanding about what the brain does. I mean, we're by no means all the way there. We're not even probably a tenth of the way there. But our understanding is rapidly growing."}, {"start": 856.34, "end": 887.57, "text": "And really, what I'd like to suggest is that if we're really interested in moving forward with AGI as fast as possible, and we believe in this system's neuroscience approach as being a useful side approach, then we should be not just passively observing what the neuroscience field is doing, but we should be actively conducting neuroscience research that might be useful for AGI questions we may have, which of course is not necessarily what's driving the majority of the neuroscience research."}, {"start": 887.57, "end": 912.32, "text": "I've argued that I believe that neuroscience is likely to have a big role to play in building AGI. Now, what specifically do I think it can bring to the table? Well, if you view system neuroscience as a kind of orthogonal source of information as compared to pure machine learning and mathematical techniques, then I think it provides two things. Firstly, it can provide direction."}, {"start": 912.32, "end": 922.03, "text": "So here by direction I mean inspiration for new types of algorithms and architectures. So I could have come up with several different examples, but the two..."}, {"start": 922.03, "end": 944.24, "text": "Pretty well-known ones are object recognition systems such as Tommaso Poggio at MIT's HMAC system, which are basically hierarchical neural networks inspired by the primate visual system. And then there have been navigation systems that have been inspired by hippocampal place cells and entorhinal grid cell findings in rodents."}, {"start": 944.62, "end": 974.46, "text": "And one could even argue that even if you go as far back as Hebbian learning and neural networks, that was sort of loosely inspired by biology as well. And the second point I think is often overlooked is validation testing. So you could have an argument with an engineer about a specific favorite algorithm of yours. So let's say it's reinforcement learning. So one could argue, is this going to be an important component for an AGI system?"}, {"start": 974.46, "end": 994.34, "text": "In the final sort of tally. And one can put forward arguments for yes and for no, but actually if we look into the brain and we find that that algorithm is implemented in the brain, then we can certainly make the case that it could be a viable, a sensible component for an AGI system."}, {"start": 994.34, "end": 1024.54, "text": "The classic kind of example of this was the finding in the late 90s that reinforcement learning and temporal difference learning were ubiquitously implemented by the brain through the dopamine system. So it does, for example, seem like that reinforcement learning is a general tool we could use as part of an AGI system. And the kind of question I sort of often argue with people who are against the systems neuroscience approach is how can it not be of net benefit to"}, {"start": 1024.54, "end": 1050.34, "text": "Add the systems neuroscience knowledge into the mix for our quest towards AGI. And just to be clear, I'm not saying we should only use neuroscience or use neuroscience as the primary lead. What I'm advocating is actually a hybrid approach where we can try and combine the best of machine learning with the best of neuroscience. And I can really sort of boil this down to sort of two points."}, {"start": 1050.34, "end": 1077.42, "text": "Where we know how to build a component already that will be useful for AGI, we just use the latest state-of-the-art algorithm. So be that reinforcement learning, Monte Carlo techniques, or hierarchical neural networks, all of which seem to be useful general algorithms. And it's where we don't know how to build a component at the moment, and I'll give some examples of this shortly, that we should do two things. Continue to push pure machine learning approaches hard."}, {"start": 1077.42, "end": 1107.92, "text": "But in parallel, also look to systems neuroscience for solutions and drive these forwards together. So what does a systems neuroscience procedure as related to AGI look like? Well, basically step one is to extract principles behind an algorithm the brain uses. Step two is creatively reimplement that in a computational model, not slavishly copying the way the brain does it. And the result should be, if we're lucky, a state-of-the-art technique and a possible AGI component."}, {"start": 1110.22, "end": 1139.36, "text": "So far, I've argued that neuroscience is an important role to play, but going towards the second part of my talk, what might we expect sensible interim goals towards AGI to look like? So some intermediate goals that I've seen proposed by some people, and a lot of people, a lot of efforts going into the moment, is on the robotic side. So the idea that to make progress with AGI, we need to have..."}, {"start": 1139.36, "end": 1155.63, "text": "Full embodied robots in the physical world. And a lot of money has been, a lot of research has been spent in this area. For example, Honda has a dedicated research team on this that spends, I think has a budget in the tens of millions of dollars a year."}, {"start": 1155.63, "end": 1182.35, "text": "for robotics and they produce impressive robots in terms of their physicality but really what we're talking about is hugely complex engineering problems of movement, server motors and mechanical engineering problems that I think are actually very difficult to solve in their own right and slightly distracting if your main interest is actually intelligence. Another quite common"}, {"start": 1182.35, "end": 1210.54, "text": "The third interim goal that I've seen a lot and some researchers are pursuing is what's sort of been termed toddler AGI. And really what we mean by that is AI control and some kind of AI controlled robot that displays qualitatively similar cognitive behaviors to a three year old. And actually from the video that Ben already showed this morning, eight months old babies are already extremely able, let alone three year olds."}, {"start": 1210.54, "end": 1237.09, "text": "What I would say about this is, and part of this definition is, what they're really talking about, advocates of this as an interim goal, is a cut-down human. With perception, actuation, linguistics, communication, social interaction, problem solving, imagination, all the things three-year-olds do effortlessly. But that's a massive breadth of capabilities required. And that equals an extremely hard problem, tough problem. So if I can pictorially represent this again."}, {"start": 1237.09, "end": 1266.24, "text": "If we, in this scale here, if on the left-hand side, we talk about where we are today with AI, and then on the right-hand side is kind of like, you know, our final goal, human-level AGI, then I would put toddler AGI somewhere way over here on, you know, NIP. You know, let's say there's 20 key breakthroughs or steps towards human-level AGI. Toddler AGI would be like step 18, something like that in my mind. And really, what I think would be more useful as an intermediate goal would be somewhere way back over there."}, {"start": 1266.24, "end": 1294.85, "text": "So what kinds of things might that be? And I spent quite a lot of time, especially over the last couple of years, thinking about what we can really boil down as being core or basic capabilities that we'd want an AGI to have that some of these more complex things could be built on top of. And so I have this diagram here to just sort of represent that. And if we think of"}, {"start": 1294.85, "end": 1321.68, "text": "Human level AGI is the entire circle and everything in it. Then what I'm proposing as being core is this kind of section here in yellow. And what I feel that most current and past AGI projects have focused on are really some of these peripheral things. So I don't know if you can read them with this typeface, but there's language here, so linguistics, logic systems."}, {"start": 1321.68, "end": 1346.06, "text": "and embodiment. So what I'm meaning there is actually robotics. And a lot of AGI projects actually focus on these what I would call more peripheral areas as the key part of their project. And I think there are more fundamental things that need to be solved and are probably blocking us from making serious progress. And some of these relate to, again, what Ben was talking about earlier on today. So I've split this into two main..."}, {"start": 1346.06, "end": 1370.98, "text": "Core areas that I think are things that we should be doing first. So difficult things that we should be trying to crack first. So one is conceptual knowledge acquisition and representation. And I'll come back to the definition of exactly what I mean by conceptual knowledge in a second. But abstract knowledge, so beyond just perceptual knowledge. And then if you have that kind of knowledge, then the ability to plan with that knowledge."}, {"start": 1370.98, "end": 1402.75, "text": "Now, in order to do planning properly, well, the reason I've lumped in prediction there is in order to plan properly in a real environment, you need to be able to predict accurately how that environment or opponent is going to react to your plans so you can adjust your plans correctly. So these two things here, and there may be other ones you can think of, but these two, I think, are more core than, say, language or robotics or logic to the intelligence problem."}, {"start": 1402.75, "end": 1424.7, "text": "My belief is actually one of the big blockers is this idea of conceptual knowledge. So let me try and be a bit more clear about what I mean by that. So knowledge in the brain, so taking some inspiration again from neuroscience, has really been talked about, broadly speaking, as three kind of levels. And these are not discrete. They're kind of fuzzy and they blow into each other."}, {"start": 1424.7, "end": 1443.6, "text": "But at the bottom level, we have perceptual information. So that's the sensory stream, all the information that comes in through our senses. And then the next level above is what I'm calling conceptual. So this is now when information has been slightly abstracted. So if we give a concrete example."}, {"start": 1443.6, "end": 1468.86, "text": "Perceptual might be what a dog looks like, physically what it looks like. But conceptual information about a dog might be things like it's a mammal, makes a good companion, it's good for guarding things, things like that. So that's the conceptual level understanding of what a dog is. And then finally there's a symbolic level where we're talking about the symbol or the word dog."}, {"start": 1468.86, "end": 1487.09, "text": "There are three kind of levels in the brain. And if we now look at where we are on the machine learning side or the AI side, we have quite good ideas of how to do two of these levels. So on the perceptual side, so dealing with the sensory stream, we actually have a number of..."}, {"start": 1487.09, "end": 1513.44, "text": "These are all algorithms and systems that can deal reasonably well, not as good as human yet, but reasonably well with distinguishing, say, dogs from cats. So there's deep belief networks, Geoffrey Hinton's deep belief networks, there's Hmax from Podio, and there's Geoff Hawkins' work on HTNs. And they're all really based around hierarchical neural networks. And they deal quite well with perceptual information."}, {"start": 1513.44, "end": 1538.06, "text": "And then at the top level, you know, we've had sort of 30, 40 years worth of research on symbolic systems. And we know pretty well how to do, you know, handle symbolic systems and logic networks, probabilistic logic networks and so forth. But what we're really missing is this kind of chasm in between. And if we could solve this chasm, this missing part that corresponds to the conceptual knowledge level in the brain, then we could solve kind of two"}, {"start": 1538.06, "end": 1566.61, "text": "Of the most pressing problems there are in AI, one of which is the symbol grounding point. We could ground our symbols in the logic networks in real concepts. And also this issue of how do we get past just perceptual information and abstract that to something that isn't directly tied to the sensory inputs and the percepts, which is something that none of these systems, although they can handle the percepts, can reach. So none of them so far have been able to generate abstract information."}, {"start": 1566.61, "end": 1592.62, "text": "Abstracted away from the actual percepts themselves. So interestingly, on the neuroscience side, we actually have quite a good candidate for how the brain does this, and how the brain creates conceptual knowledge. So it seems like there's a lot of gathering evidence, and there's more evidence that needs to be, more research that needs to be done, but there's a lot of converging evidence."}, {"start": 1592.62, "end": 1623.22, "text": "suggesting that this conceptual level, conceptual information, or you can call it semantic information, is basically constructed in the bane by the interaction between the hippocampus, which is here in blue, and higher level cortex, so association and prefrontal cortex. So association cortex is generally here in the temporal cortex and prefrontal towards the front. Now, the hippocampus is actually, just anatomically speaking,"}, {"start": 1623.22, "end": 1652.56, "text": "Extremely good candidate for doing this kind of work because it actually sits on the top of the sensory streams. So audio from your hearing, smell and touch and vision all converge at the top levels once they've been processed a lot at the hippocampus. So the hippocampus is in a kind of privileged position as far as sensory information goes to sort of tie all the multimodal sensors together."}, {"start": 1652.72, "end": 1678.08, "text": "And indeed, that's what it seems to do. So what the hippocampus is known to do already is that without your hippocampus, you're basically amnesic. So the hippocampus is critical for storing multimodal memories of your recent experiences or things you've done. So without your hippocampus, you basically have no memories. What also seems to be very clear now is that during sleep,"}, {"start": 1678.08, "end": 1705.38, "text": "And actually, not just during sleep, but also quiet moments of being awake. The hippocampus actually replays those memories, reactivates them and replays them. And during sleep, not only is the hippocampus replaying these memories, it's replaying them at orders of magnitude faster than they happen to you in real life. So up to 100 times faster. So what does this immediately suggest to a computer scientist? Well, this is an interesting..."}, {"start": 1705.38, "end": 1734.69, "text": "If you're thinking about high-level neocortex, like your prefrontal cortex, learning higher-order rules, more abstract information, then this seems like a very good teaching signal or very good source of samples for high-level neocortex to learn from. So furthermore, this whole replay system, which is called consolidation, has some other benefits too. So memories are sort of stochastically selected for replay."}, {"start": 1734.69, "end": 1761.17, "text": "And rewarded emotional or otherwise salient memories tend to get selected for replay more often than mundane memories, which is exactly what you'd want for this kind of system. So what this allows the system to do compared to sensory cortex is that sensory cortex you can think of as building up statistics of the world. So your vision is building up generalities about what it sees in the world."}, {"start": 1761.17, "end": 1787.02, "text": "But what we'd like higher order system to do is actually to be able to circumvent that and actually emphasize something important that may have happened to you but may have only happened very rarely to you but nonetheless is vital to your survival. So that seems to be a clear advantage of a system like this in that it can kind of weight the rewarded emotional otherwise important memories more heavily than run-of-the-mill memories."}, {"start": 1787.02, "end": 1817.2, "text": "The hypothesis is, and this is still work in progress, is that this could then lead to abstraction and abstract and semantic knowledge. For example, the semantic knowledge about the dog. Okay, so if we have that kind of system, what sort of milestones, concrete milestones and abilities might we see coming online? So I'll give you several examples. There's plenty you could..."}, {"start": 1817.2, "end": 1846.3, "text": "Think of yourselves and kind of add to this. But these are some of the kinds of things that I don't think are possible to do today, but maybe possible in the next few years in the sort of shortish term in terms of research time cycles. So this is why I'm turning abstract classification. There's probably a better name for it. But the example I give for this is imagine you're training a vision system to, and what you want your vision system to do is to recognize empty and full."}, {"start": 1846.3, "end": 1875.63, "text": "So the difference between a container or a cup, let's say, that's empty or that's full. Now, the traditional way you would go about training a support vector machine or any one of these kind of standard classification techniques is you'd get a bunch of examples of both. So let's say we had 1,000 examples of containers, different containers that were full with different liquids, and 1,000 containers, different ones, again, that were all empty. And you feed them in as training data to your classification."}, {"start": 1875.63, "end": 1904.83, "text": "And then how you test it is you present it with a novel container, a new container, and you ask the system, is this container empty or full? Now, interestingly, the answer is clearly in the percepts. The answer is, you know, it's in the input pixels or however you're inputting the picture. And yet, somehow, it's a slightly abstract piece of information. So it's not as simple as sort of detecting, say, the color or the shape."}, {"start": 1904.83, "end": 1926.59, "text": "There's some kind of abstract element in there. And actually, it turns out that none of the current state-of-the-art systems, so HMACs, DBNs, can do this kind of classification task. And the reason I picked something like MT4, there's loads you could obviously pick that are more abstract than that, is that that seems to be on the cusp, the boundary between perceptual knowledge and conceptual knowledge."}, {"start": 1926.59, "end": 1939.22, "text": "So I could speculate, and there's not really enough time for me to go into why I think this is a problem. I think there's two issues I can quickly go through that need to be tackled. One is the idea of building knowledge on top of other knowledge. That's something that..."}, {"start": 1939.22, "end": 1967.09, "text": "Hardly anyone has really tried to do or tackled very well. So it seems like probably the reason that these classification systems have difficulty in doing this empty full kind of classification is that really before they are trained on this, they need to know things like what a solid is and a liquid, what a container is, these kinds of issues. So probably it's sort of too high level already. The second thing is actually what I would call the symbol grounding problem rearing its head again in that..."}, {"start": 1967.09, "end": 1987.55, "text": "If you label this training data as empty and full, what is it within those data sets and those pictures of empty and full containers are you actually labeling? Which pixels are you labeling? Which parts of that image are you referring to? And of course, the classification system has no idea. So here's another example of"}, {"start": 1987.55, "end": 2009.36, "text": "Sorts of things we might be able to do if we could deal with this conceptual knowledge problem. Discovery of higher order structure. So take this number sequence, this famous number sequence, 1, 2, 3, 4, 5, 6, 7, 8, 9, 1, 0, 1, 1, 1, 2, 1, 3. And let's imagine you got this in as a number stream. Number, single digit by single digit. What is the next number? Well, it's obviously 4."}, {"start": 2009.36, "end": 2038.38, "text": "Because you can see the underlying clear, obvious underlying structure here of ascending numbers. The interesting thing about this number series is if you take it out to the limit, say a million digits, actually whatever statistical learning technique you use, which relates to Moshe's argument, however much data I give you on this number sequence, I could give you infinite data on this sequence, you will not be able to predict what the next number is. Because the actual chance of one number following another is uniform."}, {"start": 2038.38, "end": 2062.98, "text": "So the chance of a 4 following a 1 is the same as it following a 2, 3, 4, and so on, out to the limit. So it doesn't matter how much data you have, you can't actually solve this problem, which is obviously a trivial problem. So that's interesting. So statistical learning doesn't seem to be enough, or data on its own doesn't seem to be enough. Impressive results, though, it gets, as Moshe's already discussed and Google already shown."}, {"start": 2062.98, "end": 2087.82, "text": "Then there's ideas of algorithms that can build sophisticated models of their environment. So one kind of an interim master and I sort of dream about and thinking about is that which would be, I think would show clear progress in this area is some kind of algorithm or computer program that can just through observing through a raw video stream of a card table with two humans playing a card game."}, {"start": 2087.82, "end": 2118.21, "text": "Infer what the rules of the card game are and be able to beat those humans, say, after 24 hours worth of video, including if the humans just invent any game ad hoc on the spot, obviously, as long as it's a coherent game. So it's that kind of ability, I think, that is missing from what we're able to do today. And another example is in psychology, what's called transfer learning. So there's tons of examples of psychological experiments that test this."}, {"start": 2118.21, "end": 2149.28, "text": "Feature, I think, of intelligence, where basically we learn a rewarding response, let's say, in one perceptual context. We abstract the underlying rules behind why that worked. And then what that allows us to do is apply it correctly in a completely new perceptual context, one that looks completely different from the context we learned the initial response in. Of course, actually, so these are interim goals that I would be looking out for or I suggest we should be working towards."}, {"start": 2149.28, "end": 2169.62, "text": "But there are already some impressive things that have already happened. I think this is a kind of a curse of the AI field that once impressive things get done, people sort of are quite blas\u00e9 about them. But for example, MoGo was the first program just last year, actually two years ago now, first program to be a professional human player at Go."}, {"start": 2169.62, "end": 2188.85, "text": "Which after Deep Blue beat Kasparov in the late 90s was thought to be the new Holy Grail. And it did it with Monte Carlo techniques that didn't involve a lot of special case programming. And then of course, just like yesterday or a couple of days ago, there's IBM's Watson that took on some human champions at Jeopardy."}, {"start": 2188.85, "end": 2217.9, "text": "Which is obviously for the non-Americans in the audience is a quiz show in the US. And has a lot of riddles, there's questions, and it's actually a very complicated problem. And in an exhibition match, they beat two of the human world champions. So there's really some quite impressive things going on. Now, sort of related to this is, how do we measure progress towards either these interim goals and towards the overall goal?"}, {"start": 2217.9, "end": 2246.27, "text": "So I think there's two approaches to this, sensible approaches. One approach is to sort of ad hoc measure success across a sort of fairly ad hoc suite of tasks and try and hope that you pick a range of tasks that test general capabilities rather than the specific capabilities. Obviously, that's quite difficult to do, and it's quite difficult to make sure that someone can't cheat with an algorithm and special case it for each specific task."}, {"start": 2246.27, "end": 2275.86, "text": "Ideally, what we'd like is a more integrated principle measure of progress, and something that's capable of producing a graph like this, where we have year, and we have algorithms tested on this measure, and we hopefully find a nice increase in performance. That's what we'd ideally like. We're not there yet, although a colleague of mine at the Gatsby is working hard on this, Shane Legg, and actually has created, with Joel Van Ness, this idea of algorithmic IQ."}, {"start": 2275.86, "end": 2304.22, "text": "So testing the IQ of an algorithm. And they've tested it with a Monte Carlo approximation to the universal AI machine that Markus H\u00fctter invented, that J\u00fcrgen talked about yesterday, this machine called, this algorithm called AIXE. And that's not a tractable algorithm, but you can approximate it with Monte Carlo techniques. And what this graph shows is the amount of resources you allow the Monte Carlo technique to use."}, {"start": 2304.22, "end": 2332.14, "text": "In terms of context depth, and then how well it does on a 10,000 randomly sampled computational environments. And you can see this system orders these agents in a sensible way. So this looks very promising. And there's still more to be done on this. So finally, I'll just end with the theme of the conference, which is just a few predictions. So it's my prediction actually that I think"}, {"start": 2332.14, "end": 2359.71, "text": "When we come to look back on how AGI was kind of cracked, hopefully we get there, we'll find that systems neuroscience understanding will have played a part in inspiring solutions to probably more than one key component of the overall puzzle. I expect we'll see in the next sort of five years some systems with some kind of transfer learning and conceptual knowledge acquisition capabilities. I think our measurement tools will get better."}, {"start": 2359.71, "end": 2376.66, "text": "And with that kind of progress, actually, you may open up whole new avenues of doing research. So if we have a good way of specifying, fine-grained way of specifying improvement, then that opens up abilities maybe to use things like hill climbing algorithms to improve our AI techniques."}, {"start": 2377.68, "end": 2407.44, "text": "And I think once some of these interim goals have been achieved, like transfer learning, conceptual knowledge, understanding, we will have a better understanding of the overall goal of AGI and possibly have more data to deal with the safety issues that several other speakers have talked about. And I think that the overall AGI is probably 20 plus years away for human level AGI. But I think that these interim milestones will show"}, {"start": 2407.44, "end": 2415.41, "text": "Thanks for listening."}]}