Rise of AI Conference 2017 · Berlin

We May Head to an AI Winter

Danko Nikolic

At Rise of AI 2017 in Berlin, Danko Nikolic argues that the AI field risks repeating past AI winters because it has forgotten the original meaning of "strong AI." He explains that the term, coined by a philosopher in the 1960s or 70s, originally described biological intelligence, such as a dog's, as strongly intelligent because it learns from very few examples with limited resources, in contrast to today's usage of strong AI as synonymous with human-level AGI. Estimating that humans can distinguish between 10 to the power of 20 and 10 to the power of 48 things without new learning, against roughly 100,000 for contemporary AI systems, he argues that scaling AI to human-level intelligence requires resources growing far faster than linearly, making it effectively unaffordable with today's approach. He warns that over-promising based on small-scale demos, as happened with a major Japanese government AI project, is what historically triggers AI winters and could again trigger an economic crisis. He closes by describing his own startup's experiment, in which an evolved agent learned to recognize handwritten digits from only three training examples per character.

Key takeaways

  • Nikolic argues the term "strong AI" originally meant biological-style intelligence, not human-level AGI, citing a dog as an example.
  • He estimates humans can distinguish between 10^20 and 10^48 things, versus about 100,000 for today's AI systems.
  • He argues resource requirements for AI scale exponentially, not linearly, with the range of things it must recognize.
  • He cites a major Japanese government AI project that failed to deliver despite heavy funding as a cause of a past AI winter.
  • His startup's evolved agent reached about 91 percent test accuracy recognizing handwritten digits from only three training examples per digit.

Detailed notes

What strong AI originally meant

Nikolic says most people today equate strong AI with artificial general intelligence, a usage he traces to Ray Kurzweil, but the term was originally coined by a philosopher in the 1960s or 70s to mean intelligence achieved the way biological systems achieve it, regardless of human-level scope. In that original sense, a dog is strongly intelligent because it excels at tasks like navigating complex environments, recognizing patterns and catching prey, even though it cannot read or speak, while today's AI, despite lacking these biological mechanisms, is called weakly intelligent by comparison.

Why today's AI can't scale to human intelligence

As a cognitive scientist, Nikolic estimates the number of distinguishable things an adult human can recognize without new learning at between 10 to the power of 20 and 10 to the power of 48, derived from behavioral experiments rather than counting neurons or synapses. He contrasts this with roughly 100,000 distinguishable categories for the best contemporary AI systems, a gap of more than 40 orders of magnitude. He argues that resource needs for AI to distinguish more categories grow far faster than linearly, so that reaching human-level performance with current deep learning would require, by his estimate, a computer the size of a planet and more money than exists in the world.

How over-promising causes AI winters

Nikolic recalls a major Japanese government AI project, reportedly around a decade long, that received large funding but failed to deliver on its promises, contributing to a past AI winter that made "AI" an unpopular term for years. He argues winters recur when small-scale demonstrations, such as a chatbot handling a few sentences, are assumed to scale up linearly to human-level performance within a few years. Because required resources actually scale exponentially with the range of things a system must recognize, he warns that promising ever more general intelligence to investors risks not just another AI winter but an economic crisis, and that current narrow, vertical startup business models are themselves a symptom of this scaling problem.

Learning from few examples: a startup demonstration

Nikolic describes leaving academia to co-found two startups, one narrowly vertical and one aimed at technology that learns from small datasets with limited compute, inspired by how a child already carries prior expertise before ever tying a shoe. Using the MNIST handwritten-digit dataset, his team built an agent with pre-evolved expertise on digits zero through six, then trained it to distinguish the digits seven, eight and nine from only three examples each, without using the standard training set. On the 10,000-example test set, accuracy rose from 0 percent to 89 percent after one training round and to about 91 percent after further rounds, a result Nikolic says required significant compute to evolve the agent once, but very little to apply it afterward.

Quotes

The direction in which I would like to see AI develop is that it becomes so smart that we do not need big data, that we do not need big computation.

Danko Nikolic

We would need a computer of the size of a planet to do that.

Danko Nikolic

Questions answered in this talk

What did the term "strong AI" originally mean?

According to Danko Nikolic at Rise of AI 2017, strong AI was originally coined by a philosopher to describe intelligence achieved the way biological systems achieve it, such as a dog's, not human-level general intelligence as the term is commonly used today.

Why does Danko Nikolic think another AI winter is coming?

According to Danko Nikolic at Rise of AI 2017, AI winters occur when small-scale demos are wrongly assumed to scale linearly to human-level performance, while the actual resource requirements grow exponentially, a pattern he says already caused a major failed Japanese government AI project.

How did Nikolic's startup teach an AI to recognize handwritten digits from few examples?

According to Danko Nikolic at Rise of AI 2017, his team evolved an agent with prior expertise on digits zero through six, then trained it on the digits seven, eight and nine using only three examples of each, reaching about 91 percent accuracy on a 10,000-example test set.

Mentioned

a"": ""According to Danko Nikolic at Rise of AI 2017

Keep exploring

More from Danko Nikolic

More from Rise of AI Conference 2017

More on AGI & Future of Intelligence

Transcript

Auto-generated transcript, not edited. · 3,663 words

Show the full transcript

my friend question for you would be to think what do you personally wish from AI to do for you in the next year next two years five years ten years how would you like your this whole thing this whole enthusiasm to evolve so that you your life is better you're happier person and what would be the most important thing for you and I'm asking you this because I want to tell you what is the most important thing for me and what would I would like to see the direction in which I would like to see AI to develop is that it becomes so smart that we do not need big data that we do not need big computation yes I know it sounds not like crazy but I have another 15 minutes to explain ah you have seen a lot of talk here enough Nvidia chips becoming bigger more and more computational power you have seen another demo of a company that the demos were collection of the data lots of data that are needed

to train AI and so on right but if we want to go really really high we will need something we will need AI that can learn from small amount of data infants from fairly small processing power and I want to show you some of the advanced proble here know then probably here know than here yeah this type of intelligence which is not artificial with biological natural intelligence needs just one example of a car this kid can see this car you let it play with it for a little while and this kid can later recognized all kinds of course it cannot recognize big one small ones that can recognize toy cars it can recognize drawings over cars it can recognize in a schematic cars everything from one shot that's really great intelligence it's really something right and um when we talk about strong AI that's at the direction we have to go and want to be strong AI but to discuss that I would like to first say a few words about what strengths strong AI

mean because there's a confusion about this term there has been a change historical in the meaning of this term and in my opinion this change is unfortunate so it's good to clear it up early today I think most people understand strong AI is something that we run into AGI so intelligence machine that's as intelligent as human and I think the this came from the writing of Ray Kurzweil however the term comes from one philosopher who formed it I think somewhere in six days maybe 70s who was a critique of the AI at a time and who meant by date but they term more like a I in a way humans do it so not the level of human but way humanoid which means a way biological systems direct so um he would say for example that his dogs are strongly intelligent and dogs cannot read cannot speak cannot understand much of language a few words and yet they were strong in a way so how strong strongly intelligent is a dog just as considered a

problem robotics of walking of recognizing patterns of understanding the space around you and figuring out how to in a complex you know environment where there's like obstacles and things like that figuring out how to get from this point to that point to catch a prey to run away from a predator stalks to excellence are like way much better that job than any technology we have today so that's strong so fundamentally the the message behind that is that human not human break biological brains all of the brains or the living systems do things somewhat differently than what we have today in a technology so that we can learn with one trial now why is that important to know that's because we want to avoid AI winters now I winters occurred in history when big promises have been made we just heard a mention of that in a previous talk and fail right so many people okay we'll do that we'll do that we'll do that money was put in not as much as today more in

research than in the in the business but then didn't deliver this most biggest chunk of money that was poured was I think by Japanese government at some point they decided to build really powerful AI they put a lot of money I think it was a 10-year project didn't deliver pretty much anything of what they what they expected at that time right yeah I went to some bad so my second wish for AI in the future is not to see an AI winter right because it's not fun I'm I'm actually old enough to have experienced that personally when the time was that if you said I'm interested in AI it wasn't really looked good didn't look good right so you had to say I'm interested in the brain that was popular then this you could always change the topic so many people say to them now now it's for real this time it's for real AI is not gonna go away it's not going to go anywhere well hopefully that's the case but I would like

to mention a few causes of creating a winter so that we understand you know and can prevent them how do we how do we over promise something because we use intuition we think work on a small scale like you make a little demo a few sentences a little Chad bar that can kind of intelligently answer a few sentences and then you think well it works really good in the small scale we'll just scale it up in the in next two years we'll have it or the next five years which will be at the level of human um didn't happen so in a perception of pattern recognition and so on the same problem the same kind of things now what makes us believe today this is that this is not going to be a problem oh yeah we have improved a lot on speed and memory as we have just seen now and keep improving but our technology of cups concepts are basically not they are not fundamental different than what we had in 80s and

actually most of the things that we used to there have been invented by by 80s and now I wanted to see whether there is really a problem and try to do it in a scientific way in a sense of estimating the real numbers how much resource how much versus power we need to complete to match human intelligence and and to do that you don't calculate the number of the neurons and the numbers of synapses and so on you could do that but that's the wrong way of doing it because this way already makes assumptions about how the brain works but we don't really know how the brain works so what you have to do is look human behavior and what occurred and there's a I'm a cognitive scientist as well so I know how to estimate we know this and in did lots of experiences we can know how much how much how many different things can a human distinguish and not distinguish so we could you know multiply that numbers and I came to

the number of a variety the minimum minimum estimate is 10 to the power of 20 and the more realistic estimate is 10 to the power of 48 which one with 48 zeroes let how many different things a human can distinguish without learning new stuff already from the learn stuff an adult person a kid not yet but in adult birth now you look back how many different things in today's a I stick distinguish 100 thousands already really really really powerful machinery so they're 40-plus zeros missing in the scale that's that's one point yet a point even more stronger ones look how a I scales in intelligence right to scale an AI solution is number of customers that's very easy that's that's not what I'm talking about I'm talking about scaling in intelligence think this way you have one solution and I you spend one year solving things and then it works pretty well for I don't know say you recognize gestures so it works really well for 10 gestures or quite kind of well but your

plan is to have maybe 20 gestures in the future or hundred gestures let's just let's scale it scaling and intelligence now you can actually look back in the data look in our solutions today and probably everyone who has played or with with this stuff and program whoever is in machine learning here in the room already knows has feeling for that how much resources do need to distinguish two categories and images to certain level of performance then how many resources you need to distinguish ten how many resources you need to distinguish thousand or ten thousand and when you look the solutions you see that they did the number of you know Ram processing power and so rows not linearly but we're is the linear right and that's exactly why Nvidia is making so much money because people need this power that's exactly why the stock prices went up it's like exactly this problem uh-huh to reach human intelligence with this math that I've just mentioned to it just not not doable you know we would need

a computer of the size of a planet to do that I also covered in some place how much money we would need if we bought 1 gigabyte of RAM for 1 euro all the money in the world wouldn't be enough to reach this the lower limit estimate of what we what human can do as intelligent behavior with the technology that we used today with deep learning and and so on so we need to change something now what what does this have to do with AI winter but we are all fine if you are not promising too much in the terms of scaling air so if you have directly the direct advice to you from my analysis if your interpreter or if you're an investor if you have a good thing working some solution and you just need to scale it to many customers no problem if you have a good thing working but you need to scale it to working even better more intelligence more documents and more gestures recognize more herons recognize more this

more that that's where the problems begin there's where the things go exponentially we take much faster than the than the computational power much faster than the Moore's Law or something like that that's where we have to be careful if water money gets poured into prediction that in the future we'll have much more intelligent AI this could produce not only in winter but also economical crisis okay and I think this problem of scaling is also the reason why we have business models today as they are so you have startups you go vertical they go from one solution because of this problem of variety you cannot reach that broad intelligence you have to narrow down into some problem the you cannot it would be a deal if we had some platform that can scale easily as I described at the beginning without adding lots of resources without adding lots of RAM and then a huge training data set but we don't have it so you have to go vertical and then you have horizontal is reserved for

big corporations to do provide the surfaces or for the open source to provide the the tools for programming implement implementing solution but so I want to spend less my last few minutes in a telling you a little bit about how I would like to solve this problem I would like to avoid AI winter and how I would like to go and walk the path towards stronger intelligence the words intelligence that can learn only from from few examples and from from with the small we with little resources and to do that I would like to I would like to say that I have been working as a neuroscientist for many years trying to figure out from the brain what is the data brain it has that that we don't have in AI and a couple of years ago I thought I figured it out and I published lots of papers like very academics was also cool but then I decided to turn and become or go into industry and learn about that world that was completely

unknown work for me I was just a nerd in the lab now I'm a nerd in doing something else and try to learn more about business about technology but how people interpret I also wanted to get to know lots of people who could do machine learning because I didn't know I didn't I wasn't connected in this direction right and since I started co-founded two startups one going completely vertical just making one solution with existing tools not really interesting to talk about it's about psychological assessment of people and then AI telling you what people are about how to interact with them what kind of person put them in a category right something we all like to do but the other one not going completely horizontal but kind of obliques are developing a new technology that can learn on the basis of the steer and responded in this year that can learn from small data sets and something that could like this kid right the kid comes to world with lots of expertise already it's not blank

it's a table rasa like your deep learning networks actually already has a lot and then up to this age when it starts playing with car it acquires knowledge in such a way that it uses it as its expertise to learn more efficiently to have all kinds of induction biases if you like to know have it in terms of machine learning and it's a bit of a complicated theory which I don't have time to explain now you can see a we can google it of course there's lots lots of papers but what I want to tell you in the last few minutes I want to show you a result with that what we were able to implement explain a little bit how far we got and and then finally I want to show you a clip from from an animated video where we just a couple of days ago published in the video that last eight minutes and I have only four left so I'll just show you one minute from the video right so what

we did was a common power in machine learning pattern recognition this is Emma set very popular sites very well known well started and so on and you have a training data set that has 60,000 of these handwritten characters if test dataset has 10,000 of this Christmas characters and if you use this whole set you can train lots of good kind of algorithms to do near perfectly like 99 point something percent but what if you want to use just three not everything just three if you want to teach the algorithm to to distinguish seven you use three examples of silence and distinguish them from a three examples from a tenth and so on and today it's really hard to do it with with existing at all so what we did was we use this digit zeroes to six to develop expertise of the of the of the agent so it's kind of like you know what the evolution did for this kid we did with this and we developed an agent that was an expert for

four characters and his name is mr. character now but he didn't he doesn't know if I deliver it to you he doesn't doesn't do well with any characters he's just he's just able to learn characters quickly and so um here's actually I show you the results how it works actually we have we take them seven eight and nine as to train it but we use only three examples we collect them by chance so we don't do some most representative one just three by chance three of those three of those run one apart at the beginning of course it's a zero performance if we don't train it it's no surprise oops can I go back can we go one back one click back thank you oh that's not a good color do people see this line or not now I'll trace it with a laser I hope I don't burn that the walk here in here yeah so after the after the first episode it means just seeing this nine nine pictures once training ones this

one round it goes from zero to eighty-four no I'm sorry to 89% correct and then with another round to 90% and then it stays on 90 but there's a training set it's not bad because we have only three but it's just a training set so the interesting thing is actually the test the test comes with ten thousand other this is another file if you wish which completely different digits right so how does this now do with the training set right and here's the result so this one is visible I don't need laser so we get after one o'clock we get 89% e 9% and then in the next one we get to 1% and that's it right on the test so we could it so this performance is not impressive for eminence if you do it in a classical way you get to 99 percent and but here we did not do any of the optimization any of this stuff we just wanted to make it work and we were we would be happy if

it worked with with 40 percent just above just level but it worked just that with without without so what is what does that mean now here it means that we have an agent I'm at zero zero so conclusion in the video we have an agent that is able to learn any character doesn't need to be 799 it could be anything you just make it up it will just give you three examples you could probably give you just one example but in that case you have to make sure that this is really representative example here we use this random examples of three was good to cover it in in in a in a in no time you can run it on some small computer you don't need super powerful you can write down Raspberry Pi if you want um there's a need lots of resources but let's now see how many how much how many different things it can do and how much intelligence does it it depends how we counted if I this would assure

it has a really not much three right but the beautiful thing about is that it can learn any writing set how many different how many different different writing systems can you possibly have infinite more or less right a little less but for a practice practice purpose is infinite if you can learn any of them just write it in front of it and you need a few examples and then it goes and it does pretty good job so if you calculate your intelligence at that level with this with this fast learning stage then it's really huge then it then you don't have this limitation that I was talking before now here's one little thing I have to admit so I said you don't need a lot of computational power well to evolve the agent you do need we needed you know our our graphic arts got really hot to evolve it but the great thing is evolved it only once you don't have to train it for each additional problem very applied you know you could

evolve it only once and sell it as a product and people can can use it for their own purposes on on with small computation and the last thing is the video actor poiesis states that adaptive mechanisms of different speeds are organized into a hierarchy first evolution creates jeans at a painstakingly slow tempo then gene expression mechanisms act as slow adaptive masters which enable us to learn about the world by creating new mechanisms for fast adaptation next fad adaptation mechanisms change the properties of our nerve cells within seconds finally the resulting adjusted networks of nerve cells root sensory signals to muscles with the speed of lightning only then can behavior be created probably the most groundbreaking aspect of practical poetic theory is that our intelligent minds are not primarily located in the connectivity matrix of our neural networks as it has been widely held but instead in the elaborate knowledge stored in the fat adaptive mechanisms the more knowledge ourselves have on how to quickly adapt the greater capability we have to adjust in novel situation

solve problems and generally act intelligence this all the the entire video is 80 minutes it has more information but it's just a teaser and please go to youtube take a look in and Cruzer our review view account and hopefully learn something interesting so that's all what I had I'm in red please finish your talk I'm finishing thank you very much

Rise of AI 2027 · Berlin · May 2027

Be in the room for the next one.

Ten years of talks started as conversations in one curated room in Berlin. The next edition brings 300 selected leaders together in May 2027.