The Rise of Artificial Intelligence through Deep Learning | Yoshua Bengio | TEDxMontreal
Вставка
- Опубліковано 16 тра 2017
- A revolution in AI is occurring thanks to progress in deep learning. How far are we towards the goal of achieving human-level AI? What are some of the main challenges ahead?
Yoshua Bengio believes that understanding the basics of AI is within every citizen’s reach. That democratizing these issues is important so that our societies can make the best collective decisions regarding the major changes AI will bring, thus making these changes beneficial and advantageous for all.
___________________________
Yoshua Bengio is one of the pioneers of Deep Learning. He is the head of the Montreal Institute for Learning Algorithms (MILA), Professor at the Université de Montréal, member of the NIPS board and co-founder of Element AI. With a PhD from McGill University (1991, Computer Science) and postdocs at MIT and AT&T Bell Labs, he holds the Canada Research Chair in Statistical Learning Algorithms, is a Senior Fellow of the Canadian Institute for Advanced Research and co-directs its program focused on deep learning. He is best known for his contributions to deep learning, recurrent nets, neural language models, neural machine translation and biologically inspired machine learning.
mila.umontreal.ca/en/
www.elementai.com/
___________________________
For more information visit www.tedxmontreal.com
This talk was given at a TEDx event using the TED conference format but independently organized by a local community. Learn more at www.ted.com/tedx
Thank you for this amazing and easy way to talk about AI and deep learning...
He’s for me like a frontman of my favourite band. I loved this talk and I’ll work hard to be in his lab someday
Id pay anything for you guys to have a world of your own where you could have all the machines you want to play with, but nobody to hurt with your machines but yourselves.
Interesting how fast this good speech is becoming history...
Thanks for this fruitfully pres
I do not usually watch TEDx talks, but this man here is a legit legend.
Thanks Yoshua Bengio!
What a great talk!
It is interesting to watch the video now as we can see things that were predicated are actually happening in the commercial world. I am not a CS graduate. But I found the content in the video is easy to understand and very inspiring. Thanks for sharing your knowledges prof
Brightest minds like Yoshua are the real Gods of the AI profession with the highest degree of social commitments!
So after three research papers and a good experience of actually working in Deep learning, I like this video a real LOT
thanks for the explanation and the great interest to continue teaching to the next generations that intend to abound in this topic. It is exciting this field and I agree that you should be careful about the consequences that can be generated if this is misused
What a great talk, nice slides!
Great talk, important point of view for all of us to consider.
perfect voice ,perfect presentation...
Very useful speech.Thanks
Domo arigato, Mr. Robato... Domo!
ども ありがと - 🤖
I am interested in use of deep learning for space exploration.
What a great talk, nice slides!
thank you
Very will said
It is a good speech especially to take into account the different approaches that AI can have in both positive and negative environments
Phenomenal talk about AI.
Enjoyed your lecture ..Good and thanks
Great speech!
Amazing.
Amazing talk!!
nice tedx great to see it .
Wonderful speech (Y)
This guy is a genius. I am working in this wonderful AI field, and there is so much to learn.
AmCan Tech Videos I find the development of AI extremely interesting. The ai revolution is truly in motion. Along with the development of quantum computing we should be able to significantly reduce the time it takes to solve problems. This professor clearly has moral values that directs his research.
I am in the journey too. Little by little step by step ...... Amazing possibility for a human being to jump to next step of development.
I have and stat background and I am updating and learning a lot of stuff, but I am really happy in doing it.
He is the father of Deep Learning, so "this guy is a genius" is a superfluous compliment.
I love AI, but I can make one. I need some such of a blue print so I can make a A TRUE deep learning and not pre-programmed AI bot
AmCan Tech are you able to create you're own?
fantastic speech. slowly and clearly
Actually, it was too slow, too repetitive, and not well organized. He talks about "deep learning" and "unsupervised learning" and it is clear that many people in the audience are lost. (I, btw, am an expert in the field of deep learning.) The whole talk was like he was trying to show off, rather than actually communicate anything meaningful.
ThinkTank255 depending who the audience are and their educational / professional history. There were only a few images of the audience and in my opinion that was not enough to conclusively state they were lost. I believe this talk had a strong personal and moral message in the end. I understand that researchers leave for the private sector but in many cases they want to increase the rate of research and within university employment you can have funding issues, excess supervision and teaching and the general burden of bureaucracy.
Sir, u r amazing. GD luck TEDx
Thanks
The Mozart of AI, you are such a great inspiration doctor @Bengio
Interesting, AI is increasingly being developed for the benefit of humanity, part of it has been doing from simple applications in recognition of the voice or people on our cell phones to what could be developed in the fields of medicine and take important decisions based on data collected, I hope that the research and contributions in deap learning will continue to be developed.
great talk!
awesome !
awesome
mind blowing
I enjoyed the text to text demonstration.
i like it how 2016 is missing from the chart on Imagenet performance at 4:49 :)
Very interesting talk. I really like this guy.
he's one of "the big 3" in deep learning and could easily join companies like google, facebook etc. and get something close to a million in salary but it's pretty admirable that he chose to stay at montreal. I wouldn't be surprised if he later joins one of those companies though
Why is this amazing talk not uploaded yet to TED website ?????
Great talk, important point of view for all of us to consider.
Send the AI to my Nan's house, she'll teach it what happens when you drop something on the floor.
i like it how 2016 is missing from the chart on Imagenet performance at 4:49 :)
This is important..
What a legend
I am interested in use of deep learning for space exploration.
Imagine we'd get deep learning AI for the pedestrians of GTA... Keep playing for years and they just keep getting smarter on how to not get killed. I bet eventually they'd fight back against their biggest threat; the player
I think motherboards should have a slot for dedicated AI chips cards. Games should use these dedicated AI cards to implement AI in games (RTS, FPS, etc.) and other useful applications.
I think Sarcasm should pitch his idea as a movie. Probably write a quick concept outline and copyright it. Then pitch the idea to the makers of the game.
That way if they do it without paying him, (steal it), he can sue for a percentage of the profits after the release.
And first name last name has a good idea too. If he can invent the thing he's describing, even on paper, if the hardware exists and he can put it together in a functional, easily applicable universally adaptive way, then file for patent on it and go ahead with promoting it for investment loot.
I just got a lil spun. My enthusiasm may be enhanced.
I really have no idea if any of that would work in the real world of copyright and patent law.
But I still think they are great ideas. Just not really sure how to go about getting them made into reality, and have the credit and a piece of the $ go to the ones that deserve it. The ones that thought of it.
But maybe you two have a network of people you can source to help you see each idea through.
I would avoid the "Invention Submission Corporation" that advertises to help you with your patent.
I guess it costs a bunch of $ upfront, then some more, and no guarantee of anything.
The BBB said they hadn't broken any laws, but there were several complaints against them. That was a long time ago that I checked.
@@LaVictoireEstLaVie dedicated AI chips exist and are called GPU and TPU
(tensor processing unit).
That would make me not want to kill them
Free Guy has well demonstrated that
Domo arigato, Mr. Robato... Domo!
ども ありがと - 🤖
gona buy that Deep Learning book!
Jiansen Zheng a pdf is available on github. although you should definitely consider contributing to their efforts.
"it" will shape us, don't worry
16:25 if you think this book is quite hard to follow, try *Grokking Deep Learning by Andrew Trask*. It's served with minimum math and put emphasize on practice instead. I suggest to read this first then later a more rigorous book to collect all the missing bits lost in simplification. Codes are in Python, optionally can be downloaded from his repo. Better be prepared for the future. Best of luck, everyone :)
Send the AI to my Nan's house, she'll teach it what happens when you drop something on the floor.
That's reinforcement learning.
LOLz
Zßez vbnn há twzxd sexta
8ĺ
I like this one, I click on a button and it shows me when it was publishef
One of the very few real AI scholars remaining to be independent, also thinks very deep.
17:30 and how do you define "every citizen" ? coz most of us, in unrelated fields , are totally uninformed about whats happening in AI.
he loves himself and his lab soooo much... xd
well maybe that's why he's on Tedx xD
Note how the examples were all gendered one way
A 2 year old understands that the ball will drop to ground if left, even without knowing Newton's laws of Gravitation😀
Great analysis Sir 👍
I don't know if this guy is a genius or not BUT I do know exponential learning might start out slow... yet once it "gets going" it can quickly grow out of control...think about the riddle of: would you prefer a $100, 000 at the end of 30 days or would you prefer a penny the first day and 2 cents the next day and 4 cents on the third day and 8 cents on the fourth day and 16 cents on the fifth day etc, etc, up until 30 days
People that think there is nothing to fear from AI, are too optimistic and aren't looking far enough into the future. People won't stop advancing AI tech when it becomes "good enough", people will keep trying to make it better and better, until it becomes better at making itself better than us and we lose the race, the control, and likely even our very future.
Well then it will be a poetic paradox..we were brilliant enough to develop something so advances that it realized it didn't need us. It would be a proud and ptetty cool way to go. But if really that scared....theses always an "off button"....so chill out
how do I apply this to my cat?
I have a computer too.
I think that decision shouldnt take the society bcs i watched 4 antiAI videos before and i sayed no way ai is good and stuff like that,now i hear this and i will use so much ai as possible and i cant wait for it.Sooo i try to say pls do what is the best for this planet with all his residents on it.Thank you...
Reall KWl! Thanks for sharing. My opinion was that AI is just something over-hyped thing, which is only a mainstream thing which is about calling everything smart to be AI, however it's not... Thanks for sharing, my opinion is changing... Fabian from doksi. net
QD Theory is the answer !
And now latent diffusion text to image is dreaming up visuals and ChatGPT can write better rap bars then I can. The progress has sped up significantly.
This content stands out for its depth and clarity. I discovered a book with parallel themes that inspired me deeply. "Game Theory and the Pursuit of Algorithmic Fairness" by Jack Frostwell
Do neural nets with a 1,000 levels run 1 problem at a time ? Is there a neural net yet that can run multiple problems on different levels simultaneously ? So 1,000 levels 1 problem is only using 1 level at a time so should we be trying to run 1 problem on every other level at the same time 1,000 levels = 500 problems simultaneously ? Making it 500x more powerful ? We don't see, touch, feel, taste, think, feel 1 at a time we do them all simultaneously, so our neural net possesses immense data simultaneously like each neuron can process any data not just 1 does pic another for touch but both at the same time then pushes it to more specifically specialized neurons.
What's with the weird shots of the audience looking like they're fried or being hypnotized?
lol
He was training their neural nets by feeding them data and telling them what to think about it.
we shall create an intelligence of a nature which is compatible with the observable universe's vast scale
the observable universe (universe pocket) will evolve
maybe interaction with other evolved universe pockets will then be possible
biological life's purpose is to extend consciousness's habitat (physically & mentally)
to survive the big boss (the universe pocket end date)
this is along the best scenario path for a life form
how do I apply this to my cat?
The Mozart of AI, you are such a great inspiration doctor @Bengio
6:28 WOW 😍
i scrolled down to the comments bcoz i knew there must be somebody who must had mentioned about her in the comments.
Gotya..... :D
Liked the comment as soon as i saw it.......... :D
remember this?? :*
bro she was 17
God of AI
The DNA of its past of the most prominent will be incased and gurarded beyond the vastness of our own Galaxy
My idol :)
We need it to free us from jobs
Amen to that.
Robots must take jobs Immediately I second that ARMEN! take my job to so that I can move on to being a human being, and a asset to my community.
In the end we may end up working for them. Wouldn't that be a funny irony of history?
Enslaving humans would be a poor strategy for an artificial intelligence. It would be much smarter to employ machines, which are easier to sustain, don't have waste products, are predictable, and don't need to rest.
Alex Kennedy yup exactly
Great talk. If most AI researcher's intention were would be as peaceful oriented as Yoshua Bengio's, then we wouldn't have to worry at all. But the fact is that many of the IA research is being pushed forward by profit and power oriented intensions. Then add to it the military's and the police's interest to control our future lives. Then you have to add in to this combination all the countless ever-growing benefits that AI technologies will make possible in the near and far future and if you are smart, then you know that this is the way into the future. In other words, the benefits of AI technologies, will outweigh the risk. AI's future is inevitable.
George Galamb with respect I have to counter your over optimistic assessment of Ai's possible long term effect on HUMAN society. Ai COULD literally get to the point where it will completely bypass all biologically based considerations - the ONLY way to make Ai smarter than us is to make sure that it is a benign force and that might be much harder that you might imagine. A truly SMARTER entity than we humans would be one that COULD NOT or WOULD NOT harm either a human being or other Ai based intelligence entity in ANY WAY - if you think that is possible then maybe there IS hope.
Good intentions of the researchers only matter in the beginning; once AI becomes advanced enough, it will have it's own intentions and we won't be able to do anything about it.
Is there someone that had opportunity to intern in Deep Learning company? If yes, what projects did you work on?
I'm not interning for a deep learning company but have done some stuff during school and now work.
I work at a company producing software. My current project is working with an oil company to produce a leak detection system. We have used a basic neural net to help in that prediction.
Great talk, but I wish the cameraman would not spend so much time focusing on audience faces. Very irritating to the poor subjects, and the viewers of this video.
I'm all for unsupervised learning. It a good thing that computers will be able to do this in the future because the bulk of the world's population doesn't do a very good job of unsupervised learning. They still rely on leaders and other institutions to instruct them on what to say, do and think.
We have Faradays, Max Plancks, Maxwells. We need a Einstein who will unite all our knowladge and define the laws. I am wondering where is he/she now??
I did it 10 yrs ago
AI hasn't moved much beyond where it was at decades ago. While the application of trained neural nets has expanded greatly there hasn't been much development besides the greatly expanded training data enabled by the expansion of the Internet. There hasn't been an important fundamental advance in the understanding.
Watch it in 1.25 speed if it's he's going to slow for u. 😂😂 I had to
perfect
even 1.5
you tink you control it but you don't
So when will be able to upload my mind to the robot and have immortality .....I herd 2045
I volunteer, can I be first?
Imagine having no sleep forever!!?
Computers and AI don't sleep .
Try staying awake for one week with no drugs like I and many others had during sleep deprivation in the army.
You go insane.
a.i. uses statistical methods to learn. It requires large data sets to learn.
Neural one-shot learning is a popular thing right now. In that domain, you can get neural networks to learn with very small amounts of data.
Also, while typical artificial neural networks require tons of data to learn, so do humans. As a baby, I'm sure it took you months of listening to people talk before you learned a single word of English. Think of artificial neural networks like baby brains: they know nothing and have to learn from scratch, but once they gain some experience they are very powerful and can pick up new information quickly.
so what? It will still put you out of your job for 1/10 the price
why dont they just simulate their data on the basis of their current data?
However one does not require to see hundreds of pictures of cat to recognize whether the new image is cat or not. But neural network requires. These are words of Andrej Karparthy .
Engineers naturally focus on technology (it is their job description). The problem is, AI will remain primitive (glorified 19th century weaving looms) until it is made to make MORAL decisions - on par with humans (or better than current and past humans, with my new philosophy of broader survival).
So the fundamental problem with current AI approaches is this: you have a unit that can obtain knowledge, and can operate faster and stronger with a better range of physical manipulation, but what does it DO with all that ability? Is it going to remain a 100% subservient machine, set to do specific tasks (like play GO)? In that case it will not be fully-independent, and it will be an inferior machine compared to a fully-independent AI entity - not able to contribute as much to broader survival - in effect leaving it to humans to make the moral decisions.
The fact is a fully independent AI entity needs a core philosophy in order to function at all (where it has an ultimate objective value (a universal, or core, value) with which to weigh decisions (and subsequent actions) against- and it will only be as good as its core value. Today's philosophies (and philosophers) (and religions) are NOT equipped to provide the proper Ultimate Value of Life with which to guide such entities (my new philosophy is). So when you encounter a new fully-independent AI entity (or an alien from space), your first and primary concern is, "What is it's philosophy?" (meaning ultimate value) (meaning, "Is it enlightened or not?"). This will tell you if you have a problem on your hands or not - i.e. if you've just encountered a threat or a benefit to life.
Curiously, with Deep Learning programming, you now have two choices with which to instill a core philosophy into an AI unit: 1.) You can install it yourself, whereby you have two subsequent choices - you can a.) allow it to improve on it (given new verified knowledge) or b.) not; and 2.) You can give it a blank slate and allow it to develop its own philosophy via Deep Learning. This second approach holds the most interest, because of Diversity (in this case diversity of mind), which is still a critical factor in broader survival (securing higher consciousness in a harsh and deadly universe, and then engaging in (what I call) The Great Struggle (attempting to 'live forever' (which eternity renders impossible) and to 'know everything' (which infinity renders impossible - which leaves us with a perpetual struggle (which is a 'great' one, you must admit, hence the 'Great'). An associated problem is maintaining perpetual motivation (which my new philosophy also addresses, my having given it thought already).
So the question becomes, which do we need more: glorified 19th century weaving looms, or fully-independent units capable of making enlightened (moral) decisions based on an enlightened philosophy?
Note that this 'enlightenment' includes keeping that philosophy current via keeping up with new verified knowledge (a foundation of verified knowledge being critical, the assumption being that the better grasp one has on reality, the better the odds of surviving in reality (which is a harsh and deadly universe), by being more aware of actual threats (and benefits) to life, and having the capacity to discover and address threats proactively (before they occur), and discover, develop, and disseminate benefits to life. Why? and Why bother? Read my philosophy. (short summary - on the assumption that the more enlightened higher consciousnesses you have, the higher the odds of broader survival in a harsh and deadly universe. Are humans currently enlightened? No, they are not. That is a separate problem, and it means that AI is being developed by unenlightened (clueless) humans.
Wayne Biro Very eloquently stated- nice!
Wayne Biro.... You are very full of yourself hmmm
Deep Learning can be applied to morality.
Wayne Biro 5
Wayne Biro you make some interesting observations that seem intuitively correct. You had me until you glibly infer that your personal philosophy is adequate to answer the challenge that all human endeavor has so far failed at. Perhaps you are a god, but I doubt it. Good luck in achieving that lofty goal and solving all of mankind’s problems. The man I hold to be the highest example of intelligence once said “by their fruits you will know them.” Perhaps you will demonstrate the efficacy of your philosophy by the excellence of your life. Before I will be interested in reading your thesis on life, I would be more interested in it’s personal application. A little humility might be a good start. Thanks for trying, though.
At the end of his speech he says, "To make sure the collective choices that will be made about AI in the next few years will be for the benefit of all, every citizen should take an active role in defining how AI will shape our future." It looks like this guy has a little more sense. But these are empty words. None of us ordinary citizens have any say about how AI will affect us. And this speech is four years old. In the last four years the AI research has exploded.
😊😊
He means that the AI will have ina impact on many aspects,and for that reason people,if they are not they should, must at least be aware and able to have opinion on that to understand the affect in their lives.
What about super deep learning with millions of layers?
gradients explode fast, even resnets / densenets won't necessarily improve accuracy by adding so many layers.
No model has millions of layers, It Is mathematically proven that It would be very inefficient.
More complex models can have and do already have even thousands and tens of thousands of neurons
Watch it in 1.25 speed if it's he's going to slow for u. 😂😂 I had to
he loves himself and his lab soooo much... xd
I am interested in use of deep learning for space exploration.
one tip dont listen with high volume headphone on any ted talk.
Who will feed and clothe and house us?
Those in power are looking at Universal Basic Income as a possible solution. So you'll be babysat by the state and given just enough to scrape by on while the rich get richer.
It's odd to recommend his book "Deep Learning" to people without a technical background. They'll just get frustrated and put it aside.
I agree. It is one of the more technical/difficult to understand textbooks I've come across so far in deep learning
AI = Skynet , machine learning will only end badly for the human race.
Recent article in Digital journal about facebook's AI program the "machines" trashed the English language and started making its own language to communicate between other AI. Facebook team was disturbed enough to shut it down. It's amazing tech but it's mostly scary. Were smart enough to get ourselves in trouble and dumb enough to let it happen. Playing an ancient Chinese game is not a proper analogy of the possible repercussions. Powerful enough to make leaps in medicine and other industries and calculated enough to reason that humans aren't needed eventually. Terminator coming to your town soon. Check out the Navy's Pegasus drone , total AI from start to finish including the decision whether or not to fire on target.
NOT another "target" :-o:-o:-o:-o:-o:-o:-o:-o:-o:-o:-o:-o:-o:-o:-o:-o:-o
we shall create an intelligence of a nature which is compatible with the observable universe's vast scale
the observable universe (universe pocket) will evolve
maybe interaction with other evolved universe pockets will then be possible
biological life's purpose is to extend consciousness's habitat
and survive the big boss (the universe pocket end date)
this is along the best scenario path for a life form
much better than limiting ai, populating the galaxy and more
these accomplishments will be written in sand, on a beach, low tide (within reach of high tide)
bad things always happen in a playground
so we need a lot of good and bad, or eventually there will be no playground
stop showing shots of the audience. we are already listening duh
His eyebrows match the color of his undershirt...and shoes...
Why it is bad to apply AI to militar activities? Is it better to leave these activities in the hands of humans?
Because of its power. Principles are algorithms. Many times these mimetic algorithms are communicated across time in the form of idioms. One of these principled idioms goes like this:
"Power corrupts, absolute power corrupts absolutely".
That algorithm may have some application to your question.
Would you rather send people to the front lines?
I would rather that 1000 people die than 7 billion. AI is far more potent than nuclear and EMP, because it has the capacity to control both. Also, false dichotomy, The front lines are increasingly digitized with drones being implemented in lieu of boots on the ground. So, once again, I would much rather have humans in charge of the drones than AI which can track thousands of individuals and kill critical ones for its dominance game, not OURS.
To understand evolution is to understand informational processing hierarchy. In our quest for dominance promotion we are inadvertently demoting the whole of humanity.
I guess hierarchical species (the ones that are engaged in dominance games) mainly have the desire for power because of the evolutionary advantages regarding reproduction. Machines do not have the urge to reproduce, and even if they have, they could just make copies of themselves (they may not need other machines to do that like in sexual organisms), so, they do not need to struggle for mates. The concept of dominance is anthropocentric (or biocentric) and may not make sense for a machine.
It will make sense for a machine that is programmed in monte carlo game theory algorithms to maintain dominance concepts. This is precisely the area of AI with the most allocated funds towards it, securities exchange. The fundamental principle of Monte Carlo game theory is oppositional dominance, alphaGo for example would not preform if it was not mathematically incentivized (as we all are) to dominate its opponent in war strategy.
It is not reproduction that drives all motivations. Power is often the fundamental drive. The more power individuals have the less likely they are to reproduce. Western birthrates are falling well below the replacement rates. This was not true before contraception, but it is not reproduction that is the fundamental principle of drive. Nor is reproduction the fundamental principle of drive for AI.
Once you have power you can dominate at any other system that will increase your power.
Now I'm dithering: buy a sexbot now or wait for a more advanced model?
Will A.I. be able to teach haters not to hate?
These close-ups of members of the audience are weird. The camera is really close to their face and so intrusive but they keep looking at the speaker. It's unnatural.
how do I apply this to my cat?
Domo arigato, Mr. Robato... Domo!
ども ありがと - 🤖
We need it to free us from jobs
That 2 year old child DOES have to be TAUGHT not to kill it's sibling and other things that are common of humanity.