Artificial intelligence pioneer says we need to start over
axios.com
axios.com
The history of AI goes in cycles. Someone has a good idea which solves some problems, followed by "strong AI Real Soon Now" enthusiasm, followed by that idea hitting its ceiling. AI has been through search, backtracking, the General Problem Solver, hill-climbing, and expert systems. Each was overhyped at the time, and each hit its ceiling.
The big difference this time is that the ceiling with machine learning is high enough for large-scale profitable applications. That wasn't the case with the previous rounds. AI used to be a dinky field - about 30-50 people each at Stanford, CMU, and MIT, plus a few tiny groups elsewhere. Now it's a huge field with big companies and big profits. That makes it self-sustaining.
Hinton has a point in that we're missing something. Back-propagation is an extremely inefficient method, especially since the slower you do it, the better it seems to work. More generally, most of machine learning is "turn the problem into an optimization problem and bang on it really hard with lots of compute power". This works for a useful class of problems. But it has limits.
How long until the next big idea? The last "AI winter", after expert systems, was 15 years.
My company (fortune 500) depends on expert systems, some even date back to the 80's AI boom. Doing what we do, at the scale we do it, would simply not be possible without them.
Lots of companies make a lot of money with expert systems, but they have no incentive to fund AI research since expert systems have hit a clear capability ceiling.
I'm not sure I follow the reasoning here. Is the suggestion that AI cannot do better than expert systems?
My point is, they were and are profitable, but clearly that didn't keep winter at bay.
I have previously built an expert system - a system of codified expert human knowledge built into a system of executable, interrelated rules. Rule engines are the execution engines of expert knowledge in an expert system. The RETE algorithm is another good place to start.
https://en.m.wikipedia.org/wiki/Business_rules_engine https://en.m.wikipedia.org/wiki/Inference_engine https://en.m.wikipedia.org/wiki/Rete_algorithm
When I was first learning, I downloaded and played with CLIPS, a pretty cool old expert system tool built by NASA. https://en.m.wikipedia.org/wiki/CLIPS
That each was over-hyped I can buy. I'm less convinced of the latter part of that statement. Take ANN's. You would have thought that they had hit their ceiling at one time as well, but eventually the confluence of more compute power, more data, and better algorithms resulted in the recent resurgence. I don't see any reason to think that some of the other existing AI techniques might not also experience an explosion in utility, if combined with massive increases in compute power, more data, and better algorithms as well.
The point is, progress isn't linear but exponential: each new technique multiplies our capabilities.
This is what I believe as well. I've been doing some work lately on some older approaches to automated abductive inference, with an eye towards doing some of this kind of hybridization. No results to report yet, but I do think there's some meat on this bone.
The mindset that I feel is lacking in the ML/AI community is that we should be creating systems that can be augmented and improved using AI, the product itself should stand alone without using any AI at all. This is similar to how JIT compilation can improve the machine code as the program runs. Part of how our own brains work is that the cerebral cortex doesn't really have any direct control over the body at all, it can only watch and identify patterns and project intentions to the the rest of the brain/body. This is a passive structure initially, it observes both the environment and how the lower structure interact with that environment. Only after it's observed and learned the "commands" can it begin to influence how the body moves around and interacts with the world. You could look at the CLI terminal as an example, on it's own I can interact and use it perfectly fine. But lets say I added some patterning matching and prediction AI that could make really good suggestions about what command I will type next, at some point you could promote these suggestions so that they always execute when the AI predicts you would use it. This is exactly how the brain works.
Most of my ideas of applied AI revolve around improving our HID interfaces to computers themselves. For example, I'd like to reduce our chances of developing RSI related to keyboard/mouse/trackpad usage and I feel that chorded keyboards are more ergonomic. But chorded keyboards are slower to use and harder to learn. The mapping of the chords is where I plan to apply some AI, initially I'm going to randomize how chords are bound to character inputs, these random values will use many fingers to input 1 character. As the user learns and uses the chords to input characters the AI can identify common patterns and promote those patterns into chords that use less and less fingers to input, thus improving the input speed potential of the user. The next place to use AI in this project would be autocorrection and using the user's input context(program, filetype, time, surrounding text) to continue improving these bindings and making completion suggestions.
Numenta, Vicarious, NuPic, HTM, CLA, etc.
Jeff Hawkins has the right intuition and the right philosophy. Some of us have had similar ideas for several decades. Certainly, we all agree that AI systems of the future will be hierarchical (it’s the very idea of deep learning) and will use temporal prediction.
But the difficulty is to instantiate these concepts and reduce them to practice. Another difficulty is grounding them on sound mathematical principles (is this algorithm minimizing an objective function?).
I think Jeff Hawkins, Dileep George and others greatly underestimated the difficulty of reducing these conceptual ideas to practice.
As far as I can tell, HTM has not been demonstrated to get anywhere close to state of the art on any serious task.
I still think [Vicarious] is mostly hype. There is no public information about the underlying technology. The principals don’t have a particularly good track record of success. And the only demo is way behind what you can do with “plain vanilla” convolutional nets (see this Google blog post and this ICLR 2014 paper and this video of the ICLR talk).
HTM, NuPIC, and Numenta received a lot more publicity than they deserved because of the Internet millionaire / Silocon Valley celebrity status of Jeff Hawkins.
But I haven’t seen any result that would give substance to the hype.
Don’t get fooled by people who claim to have a solution to Artificial General Intelligence, who claim to have AI systems that work “just like the human brain”, or who claim to have figured out how the brain works (well, except if it’s Geoff Hinton making the claim). Ask them what error rate they get on MNIST or ImageNet.
---- Reference:
Oddly enough neural nets languished in the frozen tundra for 50+ years if you take it back to perceptrons (mid 60's)[1]. Misnky and Papert's take down of perceptrons set the field back years. If you want to be more strict then it's been 31 years since Rummelhart and McClelland published their book[2], based on their earlier work that Hinton co-authored. They laid out multilayer NN with back-propagation. I got hold of the book in the UK in '87 and used it to code a multi-layer back-propagation network. It was a machine that would read out aloud given some text, even for words it had not seen before (good luck with through, though, and trough.) The problems were data and processing speed. I had to hand code features and wait for overnight training runs on VAXen and later a Sun-3 clocked at 16.67MHz[3]. It took big data an brute force to get the glimpses or magic that we see now. The fundamentals have been around for a long time.
It may be the case that one of the many approaches that fell by the wayside will make a comeback a al NN. NN added more layers and tweaked backprop, plus data and speed. It may be the expert systems will rejuvenate as something "new" with machines writing their own rules and somehow getting over showstoppers like the unmanageability of a large rule bases. It's probably going to be some combination of things that we already know about, along with yet more data, and processing power that will get us to the next level. The odds are that whatever that beast is will be self-training and regulating in an analogous way to back-prop but with logic thrown in. Good luck to us all for inventing such things, and for trying to understand what they are really doing.
[1]https://en.wikipedia.org/wiki/Perceptron [2]https://mitpress.mit.edu/books/parallel-distributed-processi... [3]https://en.wikipedia.org/wiki/Sun-3
It didn't go well. I happened to mention that I thought parallel programming would be a big help speeding up neural nets, and that I'd toured such a multi-processor computer in Alberta. He blew up. I never got to explain what I'd done with neural nets. He tossed me out of his office summarily and wouldn't listen to another word I had to say - somehow he'd gotten the idea in his head that I was a salesman for the company whose machine I'd seen, which machine was too weird to be considered anyway, and absolutely nothing I could say would persuade him that I wasn't a salesman. So the nets I'd created and methods other than back-propagation never got mentioned. Weirdest meeting of my life, and the end of my desire to do anything with neural nets, since other academic departments were even more violently rejecting of the idea back then.
I went to a talk from people who had built more scalable training algorithms for Restricted Boltzmann Mahcines (which Hinton was a pioneer in) a few months ago and asked them why they had chosen to use RBMs. Besides the waffly answers about how generative models were more interpretable than discriminative models, the real answer seems to have been that this was a research program that had started when RBMs were still in vogue, and their funding was tied to it, and once it ran out they were going to switch to investigating Generative Adversarial Networks like everyone else.
I don't want to knock machine learning. There's a very huge set of very interesting problems that it does solve that were basically intractable till now and we haven't even scratched the surface of applications for. But the existing techniques have understandable restrictions to the scope of soluble problem.
I would also not be surprised if the best models included current ml strategies in some way, for example, in supervisor threads or initial processing of complicated input data into semantic-ish vector valued forms.
I think there's a chance it could in principle. Or at least, I think it could yield something close. If human intelligence is largely rooted in a neural network with billions of neurons and trillions of synapses, then it seems reasonable to think that, in principle, a sufficiently deep/wide ANN could show "intelligence".
But here's the rub. Doing that would be silly. We already know how to make a neural network with human like intelligence using billions of neurons and trillions of synapses... it just takes a man, a woman, and about 9 months. :-)
Also consider the power consumption of an ANN of that scale, given the current state of things.
So yeah, to achieve anything truly AGI like, which doesn't need it's own nuclear power plant, I suspect we're going to need other approaches in addition to today's ML/DL stuff.
Supervised learning
You can judge the output of your network against ground truth. You say that's a cat? Nope, it's a dog! And then slightly adjust your network so it's less likely to give that wrong answer in the future. How exactly you adjust the network is what backpropagation describes (in combination with something called a learning rate).
Unsupervised learning
You need to learn without someone telling you what the answer is. Most of the time for biological intelligence there isn't an oracle describing the truth at every moment of life to judge actions/decisions against. If you don't have someone telling you you've made a mistake, how can you know when to adjust your network? And if you don't know what the truth is, exactly how to adjust the network becomes tricky.
Somehow biological brains work without that oracle, and there are lots of ideas about how it does that. But right now for artificial intelligence none of those ideas has been shown to work so amazingly well that it has taken off like backprop has in the supervised learning world.
Hinton wants to find that amazing algorithm.
There is: Death, without reproducing. If an organism dies quickly but left children, they were biologically successful. If you didn't, then your evolutionary branch goes away.
I'm not sure this can be simulated, but it might be worth trying. Biological brains are the product of a billion years of the above strategy, so hopefully it's feasible to do the same thing.
Don't people already do that? Well, no. The strategy implies some sort of physics simulation with general rules, not a specific ruleset tuned to your domain.
> There is: Death, without reproducing.
The problem is that it's not describing the truth at every moment of life, just the total approximate fitness of that lifeform by the end of its lifetime. It would be time consuming and cost-inefficient to simulate many possible configurations of actors until the end of their lifetimes.
edit: This comment isn't a joke, I mean this seriously. Modeling a system after ourselves may not be the best idea if we're not the best possible system.
We're biased towards individualism and the illusion of personal intelligence making clever decisions using its own resources, but in fact our real intelligence is collective.
Individually, most humans are a little bit smarter than other animals, but not much.
Humans are also excellent mimics. We learn by copying, and a lot of what we consider intelligence is copied behaviour and copied beliefs - often applied unconsciously.
So we've created systems - spoken language, writing, science, technology - that can externalise, collect, and preserve the best of our learning, so the species as a whole, including average humans, can get the benefit of it.
This works fine as far as it does, but our political and economic systems are still founded on animal logic (i.e. personal advantage no matter the cumulative long-term cost) instead of conscious collective intelligence amplification and species optimisation.
Based on this, AI won't get far unless someone invents an analogous way to externalise and abstract learning, so different AI projects can bootstrap each other, and there are common systems for abstraction and generalisation.
Unsupervised learning isn't going to be all that interesting unless it can do that. It'll be a fun toy for certain classes of problems, but it won't be anywhere close to an AGI.
Although I do tend to think that given enough time (some millions of years), other intelligence will evolve on earth, from parrots or elephants or dolphins or octipuses or something. Though there's no evidence of a brain growth spurt that led to us, in other species; and it's not settled why we had it.
If the output of your networks performs well enough when it confronts the real world, you get to reproduce before you die, and some of the essence of your network gets to live for another generation.
If not, you die, and your network doesn't get to contribute to the next generation.
>>Most of the time for biological intelligence there isn't an oracle describing the truth at every moment of life
This is false for most of learning of human intelligence. Most of schooled human intelligence looks like supervised learning.
Innate intelligence and other animal intelligence looks like pre-trained models and some reinforcement learning.
Children learn the rules of grammar long before they realize there are any rules at all. Tenses and plurals, etc.
You would be surprised how much infants learn from listening around them and how long it takes to learn that language, and how a brain interprets language was formed through so many years of evolution to be highly tuned to learn language fast.
I understood your previous comments as diminishing the sheer amount of unsupervised learning we do as humans.
The list is almost endless. In all of those cases, if there is any supervision it is minimal.
Starting to suck when they feel something on their lips _is_ a reflex though, it isn't learnt.
Clearly there is stuff pre-programmed in the DNA and some animals e.g. are able to hold balance right after birth. But apparently most of this stuff in case of humans is instructions how to build a learning system, rather then encoding the complexity of the world in a strand of DNA. I grant you, the evolution did not have the time to encode ways to handle lego bricks and yet 2yr olds spontaneously acquire ability to manipulate these objects way beyond what any contemporary robot can do.
I'd argue that we see clear examples of such in animals that do require those for immediate survival. We ought to think about it as a consequence of natural selection rather than a 'necessity' that is somehow intelligently pre-programmed for survival. It's far more likely that path to early survival traits were mutational (accidental) rather than supervised.
I think you are severely underestimating humans, and probably animals too.
People have multiple modes of learning. You can try to get an A in school ("optimize" for the objective function), but you also might be skeptical of what the teacher tells you. Likewise you might be skeptical of what you read in a book.
You can also correct the teacher based on common sense and what you already know.
Currently, a computer can do no such thing. I think he's right that there is a fundamental difference.
Humans also learn from extremely small amounts of data. (Sometimes they learn badly, but they can change their minds, or not if they still manage to navigate the world.)
What does this mean? That 18 years of schooling is a "small amount of data"? Do you think you show 2 pictures of an airplane to a child and they know what an airplane is? Do you even have kids?
You appear to be thinking of just facts and figures. AI is not that.
5-year old humans have picked up a lot. For instance, I knew about the dangers that hot surfaces, sharp objects, angry cats, and big sisters represented.
I had learned to walk, run, swim, ride a bicycle, and catch a ball.
Is it possible that your conscious self is the oracle that provides feedback to subconscious networks?
I suppose an animal may have a need - it feel hungry. If it can satisfy that need then it should reinforce whatever it did to get not-hungry. I suppose that doesn't require conscious awareness. But at some level of complexity we have to be able to plan more complex actions (hunting, tracking) and be able to self-evaluate performance on the sub-tasks used to get food.
I think the origin is communication and memory. When a being (X) signals to another being (Y) the origins of its pain (signal of increased death probability) and X and Y can remember this cause, they can begin to teach.
Long before language, babies spend hours every day just interacting with the physical environment, which is "true": the toy is here, not there. The body weighs a certain amount. If you move the head this way, you'll roll over. If you move it that way, no rolling. Etc.
And parents provide a different kind of guidance. We react differently to different sounds, actions, movements of child. Even if these are not "truth" philosophically, they represent for the child emotional truth.
I'm just a guy who reads things about AI on the Internet, but I have to say it sometimes seems to me like AI folks have funny ideas of what "intelligence" is, and how it develops. The description of "unsupervised" learning above doesn't comport with any human experience I recognize. IME, humans receive constant immediate feedback on everything we do in life.
I wish I'd heard of the conference, dammit!
Mr Hinton was somewhat vague and brief in his statements here, though. What do you think the crisis point is that he's reached exactly? (It must be sure enough for the man who introduced the paradigm in the first place to have caught on, yet the change of thought process unclear enough for him to recognize that it will be somebody newer who will have to introduce something else)
This may be where my limited knowledge is my handicap, but I didn't think you would need an AGI in order to achieve unsupervised learning. There would still be an input and an output as there is now -- but instead of massive datasets used in training the models could simple experience the problem in real time, adjusting and learning based upon new results. Of course unsupervised learning models could contribute greatly to the pursuit of AGI, but I didn't think AGI predicate unsupervised learning models (which is what I am taking from the article -- or more like article abstract -- here).
Word2vec, a variant of which is Transform, Google's new machine translation framework which beats the pants off every other approach, to name one. The system it is replacing is also unsupervised.
Unfortunately while I think many of Kuhn's observations are interesting, I'm confused what you think scientists might learn from them. After all, the work in question is science, not philosophy of science, and frankly it seems a lot of time like the latter gets in the way of work (see: Popper's radical empiricism). Poetic? Sure. Useful? Difficult to see how.... Mostly his work seems useful for bolstering the credibility of speculative Popular Science articles (or Axios in this case).
My admiration for Hinton there lay separate from my remark about Kuhn.
I'm not quite sure where you got that I think scientists must learn from Kuhn's observations. But I will say, that while it's kept me away from my other studies (for the obvious displacement of reading one thing and not the other), I don't see the harm in indulging in such reflection. But my background isn't engineering, mathematics, or physical sciences before my more recent work. It's in humanities and the arts where multidisciplinary studies are encouraged. I suppose I'll probably always carry that with me. If nothing else, perspective is a good thing to have -- and not all thoughts, reflections, or ideas entertained need be engaged.
At any rate, the book was recommended and I try to remain open minded.
Funny you should say that. I take Kuhn's whole point to be that science is far messier than the empiricists acknowledge. What is the complication that you think Kuhn overlooks?
1980's - first decade
1990's - second decade
2000's - third decade
2010's - fourth decade
I guess, technically, the third decade of age for this paper ended in 2016, making 2017 the fourth.
But this seems like silly ways to rationalize it. Genuinely seems like a typo.
Anyway it's pretty funny in light of an intro I remembered from one of his old papers:
"It would be truly wonderful if randomly connected neural networks could turn themselves into useful computing devices by using some simple rule to modify the strength of synapses. This was the hope that lay behind the original Hebb learning rule and it is the vision that has driven neural network modelers for half a century. Initially, researchers tried simulating various rules to see what would happen. After a decade or two of messing around, researchers realized that there was a much better way to explore the space of possible learning rules: First write down an objective function [...] and then use elementary calculus to derive a learning rule that will improve the objective function." [1]
ie. backprop
So actually backprop was the solution to all that initial "messing around" with unsupervised rules. Though of course to be fair (if I understand correctly) those rules had very little to do with modern "unsupervised learning" methods (e.g. autoencoders, which still rely on backprop or similar optimization).
[1] http://www.cs.toronto.edu/~fritz/absps/hebbdot.pdf published in 2003
Evolutionary algorithms have been around for a while but haven't really taken off. Maybe the problem is that you can't get from zero to intelligence with a single fitness function.
Think about it: the "fitness function" for our single-called ancestors was not how well they could thrive in a post-industrial service economy. Nor was it how well they could thrive as hunter-gatherers in the African Savannah. It was a very, very different fitness function.
To get from zero to humans, nature had to evolve single-called organisms, then mitochondria, then basic multi-celled organisms, then animals that lived in the ocean, then amphibious ones, then land-dwelling ones, then mammals, then apes, then early humans, then modern humans.
By the way, I left out the overwhelming majority of the steps.
At each step, the evolutionary pressures on these organisms were wildly different. Each set of pressures was necessary in order for the next layer of complexity to evolve.
I think that if we want to evolve an intelligent entity, we would probably have to do it like this.
Point is that you're going to end up with a massive ecosystem of organisms if you could fully simulate an evolutionary history. It's not even clear that human level intelligence would evolve. It's only happened once in our planet's history. We might be flukes.
You might have to run the massive simulation a thousand times and somehow be able to pinpoint the human-level intelligence when it does succeed.
One would have to figure out which evolutionary pressures to apply in order to select for the desired traits. For example, at some point in the past, humans and rats had a common ancestor. That ancestor had two offspring. One offspring's descendants faced a series of evolutionary pressures that led to rats. The other's descendants faced a different series of pressures, leading to humans. The designer of a multi-stage evolutionary algorithm could choose which evolutionary pressures to apply, and when to apply them. The designer could thus shape the evolution of the program in the desired direction.
The trick is knowing what pressures to apply in order to get the desired results. How could one do that? I have no idea! But if someone can figure it out, perhaps they could mimic what nature has blindly done.
EDIT: As for the worry of ending up with a menagerie of programs, there are a couple of things that I think could help. First, the designer must be careful not to take the natural-selection analogy too far. The real world has numerous diverse environments, each with myriad niches; and in each niche, different evolutionary pressures shape organisms. This sort of algorithm need not provide all these diverse niches. It could apply a single set of pressures to all the programs in it, pushing them either in the desired direction or to extinction. Second, if the designer could identify evolutionary states that are a "dead end" and unlikely to evolve in the desired direction, they could remove the dead-end programs. Whether such dead-end states can be reliably identified is anyone's guess.
My point is that perhaps we understand evolution better than we understand how to create a mind from scratch.
I think the main problem is that we're incredibly dependent on advances in computing power to make advances in deep learning. IMHO, we've only made incremental software progress in using deep networks since the advent of convolutional networks.
I do want it to happen, though. And I think it will, in spite of whatever model wins. It's just too interesting a problem. How that takes shape, I can't dream of predicting. (eg. How soon will or understanding of the use of organics in computing advance to such a level)
It assumes that outside of this biological design space lies both lower peaks and much higher unexplored peaks.
While it is likely that somewhere in the universe are higher peaks of design schemes for intelligence, it is also likely that those peaks are difficult to reach from here (4B+ years of evolution hasn't done it yet). It's also possible that the Fermi Paradox is showing is that there is no substantially better general design schema.
So, the more successful approach would probably be to seek designs for the first Automated General Intelligence in a design space that has already demonstrably succeeded, then perhaps use that AGI to search for other better design spaces.
2) Biology has to satisfy many constraints which are not hard limitations on artificial brains, such as:
* Thermodynamics concerns, surface-area to volume ratio
* A tightly controlled operating temperature of 97-99 F
* Weight (wetware brains must float on CSF because they belong to moving animals)
* Size
Point 2, this reads as some advantages compared to our current computers.
Not to mention that, as a system, the brain is highly efficient in reusing modular structures for increasingly complex tasks.
I have a feeling when people mention "AI modelled after a brain" they consider a simplification of neuronal dynamics to be how to achieve that. But that is a very naïve way to do it. Looking at it from the system level is the more fruitful approach, IMHO.
But here's an example:
I see a stove eye is black when cold, then when I see it turn red, I touch it. Ow, it hurts. That's supervised learning. Don't touch things that are glowing red when they don't normally glow red. Now, I see an iron pole glowing red. It's not a spiral like the stove eye, but it's not normally glowing red and now it is. I'd better not touch it. I can deduce that these two objects are made out of a similar material since they're normally black or gray and now glow red. I can also deduce that something glowing red means it's hot. That's unsupervised learning.
To me, unsupervised learning is looking at a fluffy object with four legs, eyes, a nose, tail, it moves, etc. and knowing that it's some kind of animal. Creating groups of things like k-means. Supervised learning is your mom telling you that this one example is a tiger. Unsupervised learning is understanding the delta between your one labeled example and the rest of the examples in your animal group is largest when the animal is NOT a tiger.
Take self-driving cars as an example. Do you not think that any AI that will be able to do this will be essentially 'given' a huge set of knowledge before it's expected to learn anything by itself?
Maybe not. I remember (but don't have on hand) that people who studied a novel concept for 1 hour and then were tested on it for 1 hour, without seeing the result of the testing, bettered recalled the content than those who studied it for 2 hours. Granted, lots of variants of such a study should be performed before drawing strong conclusions, but it may be that back propagation isn't much used.
>I see a stove eye is black when cold, then when I see it turn red, I touch it. Ow, it hurts. That's supervised learning.
Yet most people learn to avoid glowing hot things without having to touch them. I can be told something red that isn't normally red is a danger, and I learn without ever having to touch one and be told my touching it was the wrong course of action. There is also interesting things that happen with sensation vs perception. Ever burn yourself with really hot metal that wasn't glowing? When I did, my fingers perceived extreme cold, not heat. You can also do the reverse by getting a rod made to look like it is glowing hot, chilling it until it is really cold, and hitting someone with it. They'll perceive being burned.
And yes, I'm working toward that goal.
The hardest part is finding interesting challenges for the AI to conquer.
https://medium.com/javascript-scene/how-to-build-a-neuron-ex...
I asked the author how to put it into use, how to train it? because clearly this biological neuron is not a smooth function. but I didn't get any response.
Yes back-propagation can help classify and predict many things.
But semantic information needs more research. For now the state of the art is Cyc.