Yann LeCun - The Next Step Towards Artificial Intelligence
Вставка
- Опубліковано 9 тра 2024
- Recorded July 16th 2018
Yann LeCun is the Chief AI Scientist for Facebook AI Research (FAIR), joining Facebook in December 2013. He is also a Silver Professor at New York University on a part time basis, mainly affiliated with the NYU Center for Data Science, and the Courant Institute of Mathematical Science. - Наука та технологія
First part:
00:00 - 28:22 ConvNets/Deep learning from the early days and where are we now
Second part:
28:22 We are good at perception but we need better reasoning
32:50 Reinforcement learning approach
35:27 What are we missing? (SL needs too many samples, RL needs too many trials)
37:12 How do humans and animals learn?
41:00 self-supervised (aka unsupervised) learning is the way to go (according to Geff Hinton, Yann LeCun et al :) )
44:30 could self-supervised learning lead to common sense?
45:50 optimal control approach, prediction is the essence of intelligence
Great work!
The first part of this lecture is just what I've been hoping to find: a concise, informative summary of recent advances in machine learning. Many thanks to Yann LeCun and the organizers.
It's a good presentation that seems to be the same presentation posted in November 2017 on this channel titled 'Yann LeCun - Power & Limits of Deep Learning'. Look forward to future presentations where a more significant portion of the presentation content has been updated.
Good talk. Love the final question.
Excellent
The highlighted red in the summary is the key to build a fast enough intelligent system equivalent to general intelligence
Thanks for sharing!
With 10 more years of AI, will we have progressed technology enough to make a presentation that works?
Isaac Lee or at least a bot that can edit out all his stumbles and failed video play attempts. Meanwhile surely Facebook has someone in marketing who can chop 2 minutes out of this video. Nevertheless, worth watching.
skierpage oh my God this is Oscar-worthy compared to the crap editing, presentation, and cinematography that comes out of Microsoft's AI UA-cam channel, Microsoft Research.
Both companies have shown better results from their marketing departments that show that they are capable of creating well-edited, well written, engaging, and informative content. But these presentations are almost always terribly edited, under-rehearsed, communicated with obscure symbolism instead of semantics, and have 500 views or less. Usually like... 150 or so. Sometimes 15. I hope they wonder why, and think that's a problem.
Sometimes I worry these presentations or intentionally poor adaptations of lecture to online video, because of some kind of desire to prevent people from seeing it and learning from it.
Say what you want about TED's crumbling content quality (usually because of their failure to actually film ideas that are "ideas worth sharing") they have a formula that works for that kind of content, and the world's largest companies are failing to achieve that very low bar.
I don't even think lecture adapts well to online video, but it HAS to be at least as good as a TED talk.
well, he's using Ubuntu for presentation ... :/
Osama Alraddadi but what presentation app? He says "don't upgrade to a new app version just before your presentation..." I don't think the O.S. is the problem.
Thanks a lot😀
For AIs to be on par with humans or smarter you need human-like sensors, mainly eyes and ears, speech is a good plus. The main issue is an inability to store data in a way similar to how the human brain does. The future of AI is to make them experience the world much like we do, so that they can build world models as complex or more so than our own, not make specific programs.
35:26: "So what are we missing" - The 'under-question' of the century.
I uploaded a radio show from 1948 about an "electronic brain." They reffered to creating artificial intelligence as a "Pathetic Fallacy." I think it's a great term. Perhaps gifted foresight?
The blurry GAN problem can be solved with a Split-Brain Autoencoder, and the accurate model of the world problem can be solved with FIT.
Interesting ..
Lucky M.E.
FIT stands for Fractal Information Theory. Fair warning though, the videos on it are awful.
Anirban Bandyopadhyay?
Yes!
Yes!
We know that many less evolved species are born with 'genetic' knowledge - basic behavior with imprinted 'muscle' memory coupled to complex survival processes. We loosely call this instinct. We see a baby antelope within minutes of being born stand and start walking. Clearly this super fast learning process didn't start with a blank slate. Some of this knowledge is presumably transferred via DNA or maybe electrical imprinting from the mother whilst in the womb. Innatism. I suspect that human babies are also imprinted with lot more core 'understanding' at birth than medical science currently accepts. Much like having the lid of the jigsaw puzzle box makes placing pieces so much easier and accelerates positioning and relative associations and self learning once you get to grips with the pieces.
35:32 Real talk begins
If the industry is looking to develop "Artificial Intelligence" then it may well accomplish that, and possibly at the expense of authentic intelligence, which in itself is somewhat of a linguistic paradox.
I love to see other people who like AI, I have some AI demo's on my channel, let me know if you'd like to see tutorials.
About 9 years ago I started forming a theory of dynamic cognition that I presented in 2013 at my blog.
It's a theory of intelligence based on salience. At 41:00 Yann mentions the problem of babies learning background knowledge of the world through observation.
My theory provides a theoretical framework for how this happens....in truth the answer comes from recognizing that his statement is in fact false.
Babies do more than just observe...they are a bundle of learning dimensions with vision being one seeming most active but the somatosensory sense is just as important if not more so to increased ability for the toddler to make inferences about the world.
As they go from mostly stationary to motile they than integrate a significantly larger amount more information regarding the tangential contexts to vision.
This allows a massive reduction in the data space for making accurate predictions about the world from this integrated information. Reinforcement coupled with this I assert is all one needs to produce highly generalized models and coupled with a multidimensional "salience" core coupling autonomic and emotional factors and attention allow the emergence of self awareness and emerge consciousness.
The collection of articles can be found here:
sent2null.blogspot.com/search?q=salience+theory
David Saintloth
Bookmarked your website. I'm excited to learn more.
Thank you for your work
John macarthy's paper on "Philosophical Challenges to AI" demonstrated the similar thing, i.e. we need a representation of the world with a reasoning system at the heart, well the reasoning system might be learning from reinforcement or semi-supervised learning, but I guess the challenge is to create a representation of the world, which if we talk in real world is variable, like it changes with people but if we go deep into it, like rationally we might reach the core, which I think is much related to philosophy like the person might build more things above it, but it all starts from his/her own philosophical views, well they are not concrete and might change, but it might be certain that they exist.
Well now the question is how to create the model of world.
Lucky M.E. Ohh deep...
Lucky M.E. I want to build stock market trading bot but in a different way( not just by analyzing past data) but analysis emotions of people in real time...
I'll not say its impossible, but computation needed would be a lot and also quite complex (in case you want to build a reliable model), and the worst of all we still don't know how emotions work, there is no algorithm or no reasoning system for understanding emotions explicitly, which I think might be the bottleneck in the further development of this model.
Check out Anirban Bandyopadhyay's work on Fractal Information Theory. I think it might be the missing piece.
Will our overlords have common sense? I particularly liked how a human baby understands the world by 10 months and the prediction models with feedback.
Brian Starr come on, at higher level (the way humANIMAL societies function) there is no common sense at all, not even umcommon... just the usual politico-oligarchic predators living off the herd of mental herbivores, fooled by the mindfcuking prophets (from dark age catholibanism to the nowadays popular socialist religion, esp. in its neomarxist incarnation)
budes matpicu You must be good fun at parties! Take a chill and look at some of the positives that man has achieved like the telecommunications network, and landing a rover on Mars, and this is just the beginning.
you had me for awhile, then you started talking about prediction... in reality, you cannot predict all outcomes very successfully at all.
He covered that. The latent variable z covers all the likely outcomes. You have to come up with an energy function with minimum energy for all the likely future states (the rod tips over in some direction), and high energy for all the unlikely ones (the rod leaps into the air, flies across the table, etc).
With a Bicycle GAN you could theoretically produce a collection of possible future images which collectively approach a representation of all possible futures, so you don't end up with a "blurry" average of all futures.
Sorry this video is way above my head.
Oh thank god, it gets better further on in the vid.
here's the video of the orangutan laughing at the magic trick: ua-cam.com/video/ka_IlYO3HKM/v-deo.html