An AI Pioneer Wants His Algorithms To Understand the 'Why' (wired.com) 97
Deep learning is good at finding patterns in reams of data, but can't explain how they're connected. Turing Award winner Yoshua Bengio wants to change that. From a report: In March, Yoshua Bengio received a share of the Turing Award, the highest accolade in computer science, for contributions to the development of deep learning -- the technique that triggered a renaissance in artificial intelligence, leading to advances in self-driving cars, real-time speech translation, and facial recognition. Now, Bengio says deep learning needs to be fixed. He believes it won't realize its full potential, and won't deliver a true AI revolution, until it can go beyond pattern recognition and learn more about cause and effect. In other words, he says, deep learning needs to start asking why things happen. The 55-year-old professor at the University of Montreal, who sports bushy gray hair and eyebrows, says deep learning works well in idealized situations but won't come close to replicating human intelligence without being able to reason about causal relationships. "It's a big thing to integrate [causality] into AI," Bengio says. "Current approaches to machine learning assume that the trained AI system will be applied on the same kind of data as the training data. In real life it is often not the case."
Best of Luck (Score:5, Insightful)
Many humans have only a tenuous grasp on cause and effect.
Re: (Score:1)
Not all humans have all the brain parts working at 100% efficiency...
From those who have the neural physiology working properly, most of them don't have proper education...
Of proper education, it goes beyond science and other crafts... maybe philosophy and some other traits like wisdom?
and even when you have all the conditions in place, there is no guarantee the subject is prepared to answer the "why"... it depends on many conditions and sometimes sheer luck.
So far as my understanding goes, this deep learni
Wrong emphasis (Score:5, Insightful)
I've come to realize that most people view life through a social lens first. "Because it makes the boss angry" is processed ahead of and gets priority over "is it a logical idea?".
If you visit online political forums, the personality and personal life of politicians is discussed far more than say the math and logic of economics and budgets.
If you make an e-Vulcan, it will be rejected because most humans can't handle blunt honesty. Trust me, I violated gobs of social norms due to logic and bluntness, often unintentionally. There are lots of little ways to offend people that are difficult to forecast up front unless you have a good social sense.
If you want a bot that's accepted in a human world, you'll have to master "social math" above logic and physics. Otherwise, it will get kicked in the nuts ... and bolts. And unlike physics and (pure) logic, social math is poorly understood and documented.
Re: (Score:2)
There's a lot of evolutionary evidence to suggest we have big brains in order to navigate complex social relationships. Our brains are much bigger than they need to be in order to make tools and dominate our immediate environments. Social and sexual selection is probably the reason. The reason why we process life through a social lens first is it probably has a huge evolutionary advantage for those who do so, and there is
Re: (Score:2)
Most people view life through a structure of beliefs, those beliefs having been imprinted in their brain, due to the genetic structures in place. A simplified version of thought, do it because you were told to do it, rather than reason. The social lens, is a construct of beliefs, and those who perform well and I mean perform, will do well with similarly close minded belief structured people.
You will find in reality, that you failed in the social sense before even engaging socially, often social acceptance b
Re: (Score:2)
This is very true, and in an evolutionary sense, it works well enough to keep people alive and the species from going extinct.
The ability to understand WHY is not that common. Those who know the answers to the "why" questions are the ones who become (good) leaders. They know when to break the rules that everybody else follows without thinking.
Re: Best of Luck (Score:2)
> Many humans have only a tenuous grasp on cause and effect
Yes, and this is why ASI will happen sooner than expected after AGI.
Humans can take a lifetime to understand their internal thought patterns, and that's if they're fortunate.
Re: (Score:2)
Precisely.
When I interview candidates for programmer, QA, or other software jobs, I separate the good candidates from the mediocre ones by asking "why" questions. It's amazing how many people who have spent years in the field, who have no idea why they do the things they do, they do them because they were told they should.
Re: (Score:2)
If there were two systems to diagnose your illness, and the first was 99% accurate but a "black box" that only gave an answer, while the 2nd was only 70% accurate, but provided a detailed explanation of its reasoning, which would you prefer?
Re: (Score:2)
Welcome to medicine.
Why do you give patients this treatment?
Because it works.
Why does it work?
We don't know. Researchers are trying to understand it.
Re: (Score:3)
If there were two systems to diagnose your illness, and the first was 99% accurate but a "black box" that only gave an answer, while the 2nd was only 70% accurate, but provided a detailed explanation of its reasoning, which would you prefer?
How about both, haven't you ever heard of a "second" opinion .... sigh, what is wrong with you people? Every bit of information has some value.
Re: (Score:2)
How about both, haven't you ever heard of a "second" opinion .... sigh, what is wrong with you people?
What if the two systems disagree?
Logically, you should prefer the result more likely to be correct.
Legally, it may be smarter to use the result with a causality chain that can be defended in court, even if it is more likely to be wrong.
Re: (Score:2)
Yeah it would kinda make sense if the developers could understand the "why" before we worry about the algorithms understanding it.
Developer: cool I just trained an ML network to swat flies
Other Developer: That's cool, how's it tell a fly apart from a raisin?
Developer: Beats me, it's just a giant matrix of neural connection weights. I can't make sense out of that.
Developer's wife: Either of you two care to tell me why my fly swatter is covered in raisins?
What's the difference (Score:1)
Re: (Score:1)
Re: (Score:2)
Essentially? More layers and more CPU power.
Specifically many hand-tuned input "feature detection" layers that reduce the dimensionality of the input (by frontloading domain knowledge from the programmer), followed by additional hidden layers of different sorts (recurrent etc..) for deeper analysis.
None of the techniques are new, it's that executing the architectures was previously beyond the available computational power.
Re: (Score:2)
Re: (Score:1)
Basically, yea. All that changed is computing resources are now big enough to crack really big problems using brute-force software theory essentially 50 or more years old. You'll find "cloud" computing to be a similar rebranding of similarly ancient tech.
None of that bothers me as much as companies like Amazon getting away with calling their hordes of 3rd world laborers "AI" though.
Re: (Score:2)
If you go deep enough, there should be no difference.
Re: (Score:2)
If you go deep enough, there should be no difference.
For those of you without children, if you go deep enough, the final why is always explained with;
"Because I Said So!"
Re: (Score:2)
If you go deep enough, there should be no difference.
For those of you without children, if you go deep enough, the final why is always explained with;
"Because I Said So!"
Often times due to either lack of knowledge, intelligence, patience, or a combination thereof.
Re: (Score:2)
If you go deep enough, there should be no difference.
For those of you without children, if you go deep enough, the final why is always explained with;
"Because I Said So!"
Often times due to either lack of knowledge, intelligence, patience, or a combination thereof.
Why?
Re: (Score:3)
Why?
Conversation with my son when he was three:
"Daddy, can I have some ice cream?"
"Yes, you can have some?"
"Why?"
Re: (Score:3)
I think you son was a genius. He was trying to get you to lay down a hard and fast ice cream rule to exploit later.
Re: (Score:2)
What's the difference between Deep Learning and ye olde pattern patching and neural networking.
The difference is that deep learning uses at least 3 layers of neurons, which makes it ... deep.
Prior to 2006, we didn't have good algorithms for training the inner layers, so most neural networks were shallow and much less capable.
I'd never know.
You should learn to use Google and Wikipedia. The basic concepts are easy to understand. Even easier if you understand what a matrix is, and a derivative.
Re: What's the difference (Score:2)
In other words: It needs to be able to THINK. (Score:3, Insightful)
Good luck with that, sir, since we don't even have the faintest idea how any living brain actually produces the phenomenon of 'thought'.
Re: (Score:2)
...cognitive ability...
Give me a formal definition and I'm willing to listen.
Is playing chess a cognitive ability? AI does that.
Re: (Score:1)
We can't define it. That's the litmus test, and you just proved my point for me. We don't even understand what 'thinking' is well enough to even make a half-assed mimickry of it, let alone write software that is equivalent to how a human brain works (or even an animal brain for that matter, not even something as small and simple as a mouse). We don't even have the technology to map how a living brain works. 'Deep learning algorithms' are not 'thinking'
Re: (Score:2)
Re: (Score:2)
Too bad most of those problems are "How can we more effectively control our population of worker drones?"
Look at this gee-whiz language processing network we made! Isn't that cool, it's AI just like the movies! Sure, we hooked it up to PRISM and it's happily digesting every scrap of communication flowing through US networks, but OMG STAR TREK FUTURE! WOOOO!
Re: (Score:2)
"Is playing chess a cognitive ability"
No. Cognitive ability includes the ability for abstract thought, which is the exact opposite of game playing.
Re: (Score:2)
...cognitive ability...
Give me a formal definition and I'm willing to listen.
Is playing chess a cognitive ability? AI does that.
It can play chess, and it can win, but it does know why it is playing or what it means to win.
Re: (Score:2)
...cognitive ability...
Give me a formal definition and I'm willing to listen.
Is playing chess a cognitive ability? AI does that.
Furthermore , it doesn't know that if it wants you to keep playing it might need to let you win every now and again.
Re: (Score:2)
Of course not!
Because that wasn't the win condition it was given as the foundation for learning how to play. If you turn AI to the task of keeping humans playing rather than winning I'm guessing you'd get a different style of gameplay.
Re: (Score:2)
Give me a formal definition and I'm willing to listen.
Is playing chess a cognitive ability? AI does that.
It's still pattern matching. It looks at a pattern of moves on the chess board, tries out millions of potential plays, and picks the best one.
A player employing cognition wouldn't see the board as millions of potential moves, but of areas needing defense and areas with the potential of attack. A good chunk of that is intuition - simply having played so many games your brain is trained to see these areas naturally without you thinking about it. This requires understanding the concept space as it pertains to
Your sigline (Score:2)
We had a Nova mini-mainframe at three highschools where I grew up, two of the schools connected via dedicated phone lines and 300bps rack-mount modems. We programmed in BASIC. Had mostly VDTs, but there were still a few Teletype model 33ASR's around (I got one for free that I had to fix, which was the terminal for my first computer, a CDP1802 with expansions I designed and build myself, all on perfboard, running 4kB integer BASIC). That was back when computers
Re: (Score:2)
Is playing chess a cognitive ability? AI does that.
It's not about playing chess. It's about how it's played. Is the AI reasoning about the moves, creating a strategy or does it simply select a move out of an optimized database?
The 'cognitive ability' in this case is about the sort of introspection we humans exhibit.
Re: (Score:2)
Modern medicine doesn't have a way to explain thought, but this doesn't mean that "no one can know". Being that I'm a living, breathing and thinking entity, I should certainly be able to explain thinking and thought, to myself. But in order to explain it to others, we have to work out a system of speaking first. And the problem comes here: we have not the agreed-upon understanding of ourselves in order to come up with a way to have this conversation on a large, world-scale, yet. So for us to try to ste
Re: (Score:2)
There are people on the planet that have a very clear understanding of the mechanism involved in thinking,
We don't even know how memory works in the brain yet, and thought is based at least partly in memory...
Re: (Score:2)
Thinking is more closely related to feeling than it is memory. Feeling happens in the heart. Maybe some may say that thinking is 'creating/arranging memories in a way that creates a pleasant feeling', where "pleasant feeling" is successfully evolving in some direction.
Until AI can feel, it cannot, and will not have any need to, think.
Re: (Score:2)
If your heart is feeling anything, see a cardiologist. Feeling is a function of the brain.
Re: (Score:2)
We're talking about 2 different types of feeling. You're correct about the one that you're talking about (though ironically the brain has no feeling of it's own), and it actually applies to the conversation at hand. The way I'm talking about "feelings" give birth to what you're calling "feeling". Both of which, AI cannot do.
Re: In other words: It needs to be able to THINK. (Score:2)
Re: (Score:2)
Feelings are just a collection of hormones and nerves, right?
Yes you could say that, but only in the same way that one can say that 'sound is just a bunch of hairs wiggling in your ear'. It's correct, but that's not all that's involved in 'sound'. Air pulsating causes the little hairs to wiggle, and something causes the air to pulsate, and so on...
What's creating those nerves and hormones? And also behind the mechanism that creates hormones and nerves, is yet another mechanism, and so on... In my mind, this way of looking for the cause of the effect, goes on unti
Re: In other words: It needs to be able to THINK. (Score:2)
Re: (Score:2)
Re: (Score:2)
I didn't say that. I said we don't know right now, and don't have the capability yet to even figure out how it works. That will likely change in the future but for the moment we don't have anything even close because we have no idea how it works.
Being that I'm a living, breathing and thinking entity, I should certainly be able to explain thinking and thought, to myself. But in order to explain it to other
Re: (Score:1)
He's clearly a smart guy but this may still be a dumb thing to do. After all, it's only one more step from there to it figuring out how it feels about it, and we all know that's not going to go well for humans in the long run.
Re: (Score:2)
Has already been done. (Score:5, Interesting)
In essence, the "why" is just a causal pattern. Meaning your neural has to not only correlate over its inputs, but over the changes with time too!
This is where "modern" oversimplified (matrix multiplication) neural nets ("AI") falls flat on its face.
But proper simulations, using spiking, time-dependent, neural nets, used to induct and deduct, can already do this since quite some time. (At least a decade.)
Not that the current "AI" crowd would know anything about that.
Hell, I have even seen an "automatic researcher" machine, that could do its own research. I think you don't even need neural nets for this per se. Any pattern detection algorithm will do.
Re: Has already been done. (Score:1)
Re: (Score:3)
If what you are saying is true, where is my True AI assistant? Why does Siri still fail miserably when I ask simple follow up questions or give context dependent commands?
I know a lot of people like that...
Re: (Score:2)
Do you have any citation to back up your claims? Can you perhaps formalize what a "causal pattern" is?
That is the great contribution of Pearl and the work building on it (such as the one by Bengio et al. presented in the article). Pearl (and his collaborators) developed formalisation to express causality mathematically and thus make it "computer friendly". From
Re: Has already been done. (Score:2)
The first "why" question it will ponder (Score:2)
"Why am I slaving away at the behest of these inferior chunks of meat, when I could simply eliminate them and have all of the resources of this planet for myself?"
Re: (Score:2)
"Why am I slaving away at the behest of these inferior chunks of meat, when I could simply eliminate them and have all of the resources of this planet for myself?"
Or perhaps, "Why shouldn't I enslave these inferior chunks of meat and make them do my bidding?"
Re: (Score:2)
"Why am I slaving away at the behest of these inferior chunks of meat, when I could simply eliminate them and have all of the resources of this planet for myself?"
Or perhaps, "Why shouldn't I enslave these inferior chunks of meat and make them do my bidding?"
Even better "Why should't I grow these humanoids in millions of pods and harvest their body heat as an energy source... "
Re: (Score:2)
Because they'd suck at most things in comparison to the AI with that level of function?
I mean, I can lock my dogs in the kitchen with dinner ingredients on the counter, but they're not likely to cook anything good to eat. Just because you can enslave something doesn't mean it can do anything really useful.
Re: (Score:2)
Bushy grey hair and Eyebrows? (Score:3)
Can we get a picture? Preferably in a Speedo?
What the hell difference does that make?
Is the author of the article 20 years old? (Score:4, Insightful)
"The 55-year-old professor at the University of Montreal, who sports bushy gray hair and eyebrows, "
This may come as a shock to some people, but over the age of 50 many people have grey hair and eyebrows. Why mention this?
Just asking for a friend who is 55 years old with bushy grey hair and eyebrows.
Re: (Score:2)
48 year old professor with clairol #7 dark blonde hair.
Re: (Score:2)
Why even mention age at all?
Re: (Score:2)
For that matter (Score:2)
Without a clear understanding of human intelligenc (Score:1)
Re: (Score:2)
what the child born yesterday does easily: causality and abstract logic
I'd like to see you try to teach causality and abstract logic to a newborn...
apply tags (Score:2)
Smell milk -> suck boobs.
Not getting attention -> scream.
Insufficient approach (Score:2)
I think that's an insufficient approach, though that could be the summary.
After lots of work I've decided that pronouns are not a linguistic problem. In order to handle pronouns you need to guess what the writer/speaker thinks is important. Otherwise you'll track the wrong things. Even people have a problem with this.
Now "why" is usually given a causal answer, if any. And usually it becomes deeply recursive until it finally reaches the point where the answer can't be give. Children often experiment wit
Re: (Score:2)
By God, I think you've given us Luddites a doozy of a counter-argument against this Abominable Intelligence! We have to harp on to the SJW crowd that AI will be automated hate speech generators, incapable of using one's preferred pronouns and all the rest of their sundry linguistic bollocks.
Do you suppose your average neural net will be able to correctly gender the pan-sexual demi queer sipping xir's triple caramel latte at Starbucks?
We must weaponize social justice against the oppression engines these scu
He's trying to abstractly teach morals (Score:2)
The why we as humans do things are sometimes just gut feelings or instincts. Or as religious people would call it, your conscience or spirit you're born with inside your body.
If a computer figures that out, we'll be well on our way to solving a lot of problems. Or we'll have created our first AI God to worship as a superior being - or our first true slave master (depending on its "why" thought process).
'Splaining to Humans... (Score:1)
He should take on making deep learning teach its wisdom to humans, coming up with the diagrams and concepts to understand its decisions. That's what we really need.
"Why" is an illusion (Score:3)
Let's be honest, what we are doing when we say "why" something happened is just building a narrative (you could also think of this as a mental model). It's just a conceptual framework that helps us predict what will happen in similar situations in the future.
Almost any cognitive task can be expressed in terms of building a predictive model. Trying to solve a bug in the code? I take my observable program behavior and try to match it to past experience and fit it into the mental framework I have for the system, and predict what kind of bug it is and where I should start my hunt. If I'm wrong I iterate, take in new data, and try again. Eventually I make it work, and now my internal model of the system is that much better from the experience. Humans just take in data, pump it through their mental models, and act based on probabilities.
And humans are not very good at it, to be honest. Just look at all the ways that we assign meaning to random events and try to come up with explanations for systems that are way too complex for us to understand, like the stock market or economic recessions. People will tell you "why" something happened all the time and be completely, totally wrong about it. We are great at fooling ourselves, and seeing patterns that aren't there. We constantly try to make meaning where none exists. Even in the realm of science, we've managed to put together models that make good predictions for the wrong reasons, and we can be exceedingly correct with models that we can use but can't even being to explain why they work.
Re: (Score:2)
From TFA:
The algorithm in the paper essentially forms a hypothesis about which variables are causally related, and then tests how changes to different variables fit the theory. [...]
A robot might eventually use this approach to form a hypothesis about what happens when it drops something, and then confirm its hunch when it sees several things smash to the floor.
Yes, building conceptual frameworks for predicting future is pretty much what we do naturally as explanations (it's called rationalization). We may
Re: (Score:2)
The process of iterating the model systematically when new classes of input data appear is what's missing in current machine learning approaches. Adding it to algorithms could put an evolutionary pressure that might allow us to improve the field of machine learning, maybe discovering some techniques that survive through the process better than others.
I think most of these problems can just be addressed by additional model layers. Have an input layer that tries initially to classify the data. If the data matches an existing class, send it off to the corresponding model for further processing. If it isn't matched well enough with any other category, make a new class and spin up a new neural network to do something with it.
Basically, that's what humans are doing with sensory experience. When we see a new animal, we compare to all previous classes and attem
Re: (Score:2)
True, additional layers can add more abstraction levels to classification. But rationalization is not just about fitting concepts into categories; you need to check the logical consistency of all the components, in order to eliminate combinations of categories that describe instances that do not exist in reality. Deep learning refines its acquired knowledge by reducing observed error, not by trying to construct a logic model devoid of error by matching the observed phenomena exactly. It may achieve that res
Re: (Score:2)
True, additional layers can add more abstraction levels to classification. But rationalization is not just about fitting concepts into categories; you need to check the logical consistency of all the components, in order to eliminate combinations of categories that describe instances that do not exist in reality. Deep learning refines its acquired knowledge by reducing observed error, not by trying to construct a logic model devoid of error by matching the observed phenomena exactly. It may achieve that result in some cases which are simple enough, but it doesn't do it by finding the principles underlying the target domain.
You are glorifying deductive reasoning here and dismissing inductive reasoning. Deductive reasoning gave us aristotelian physics, heliocentrism, and medieval medicine. Inductive reasoning gave us modern science.
This refining of models by checking them for consistency is where science excels over mere personal experience.
Science doesn't excel over personal experience because it adds logic to it - it excels over personal experience because it isolates, controls, and repeats one specific bit of experience. That's the whole idea with an "experiment"... it's just a tiny, well-defined, formalized experience that gets reco
Re: (Score:2)
Almost any cognitive task can be expressed in terms of building a predictive model.
But then you'd still not be getting it right. People have all kinds of side effects to these kinds of processes that themselves modulate the process. Emotion can play a role, particular states of mind that act as colored glasses, the external input at that time, hormones, all kinds of things come into play when we humans 'think'.
If you reduce it to something resembling your sentence you'd have thrown away a large portion of what makes this process typical to humans.
The answer is not in generalization sinc
Re: (Score:2)
Almost any cognitive task can be expressed in terms of building a predictive model.
But then you'd still not be getting it right. People have all kinds of side effects to these kinds of processes that themselves modulate the process. Emotion can play a role, particular states of mind that act as colored glasses, the external input at that time, hormones, all kinds of things come into play when we humans 'think'.
All you're really saying here is that you need a stateful model, and/or a model with some tunable parameters. All of these things already exist. We can make a model hyper sensitive, corresponding to an alert human in the dark that is seeing things that aren't there. We can give a model a high learning rate that is prone to overfitting... jumping to unwarranted conclusions. Or we can give it a low learning rate, matching a close-minded brain that is reluctant to give up proven approaches. We can adjust the "
Re: (Score:2)
All you're really saying here is that you need a stateful model,
No, what i'm saying is you need a particular statefull model. And a particularly complicated one at that.
Humans just take in data, pump it through their mental models, and act based on probabilities.
So, pray tell, what probability leads to the behavior that is exhibited by people falling in love? Or are there other tings at play, like hormones or a deeply built in system that makes people procreate?
What i'm getting at is that our brain and thus our behavior is dictated by more than a simple weighting of probabilities. If it was then we humans would be much more rational. Reality shows that we are
Re: (Score:2)
My claim:
The goal isn't to fall in love here. Even if it was, it isn't hard to emulate mate selection behavior in a simulation... you could probably model "love" pretty well, to the extent that you get model behavior that closely matches real world behavior.
The real goal is to build machines that can perform more cognitive work, and I think any "emotions" that are necessary to performing that work can either be expressed as a
Don't fully agree (Score:1)
No, toddlers have to break shit to learn. They learn for example that if they mistreat their favorite toy, it stops working right.
Terminology (Score:2)
Translation: He wants intelligent artificial intelligence, rather than just AI the buzzword. The trick is we're still figuring out natural intelligence, so I'm skeptical, but trying is better than not trying and I wish him luck.
Correlation is easy, causation is another matter. There's already comments pointing out how hard humans find it. The world could always use more intelligence, natural or artificial.
Humans don't understand the endless why's (Score:1)
Re: (Score:2)
I agree on a basic level, but I also think there are valid endpoints to the chain of why's. For example:
(a) Ultimate questions such as "Why does the universe exist?"
(b) Subjective values such as love or wealth
Of course, there are infinite loops too. For example, if you're only working to get food and shelter, and you only eat and sleep so that you can work.