Not sure if you guys gives a damn but if you're bored like me during the covid times then you can watch pretty much all of the latest movies on InstaFlixxer. I've been watching with my girlfriend for the last couple of weeks xD
modelling -> finding out /predicting -> Basically finding out from data a model density -> here it means probability density function -> i.e. probability distribution of data/thing to be modelled.
About 17 minutes and I had to stop listening because I felt like I had lost about a standard deviation IQ. Hasn't this guy ever heard of Solomonoff induction? Hasn't he ever talked to Shane Legg? The intrinsic motivation is lossless compression and if the agent is active the decision theoretic utility determines the explore exploit tradeoff as in AIXI. If passive it just compresses whatever it's given as data.
@@theJACKATIC I listened to the rest and he did, finally, bring in compression as one would expect of someone with his background. And it does appear important. His presentation threw me off. At a meta level, he really should start with the "high level coding" of his presentation: Describe the space in terms of AIXI's unification of Solomonoff Induction and Sequential Decision Theory before breaking down into his 2x2 taxonomy. That way it would be clear that "unsupervised learning" is simply lossless compression toward Solomonoff Induction's use of the KC program's "latent representations". He appears to have his head so far into the techniques of lossless compression that he elides the "top down" definition of AGI as the start of his "high level".
@@jabowery some persons are not so advanced in this field and starting from the known to the unknown is a great technique in passing knowledge. Great presentation🙋
@@coolmechelugwu7305 Solomonoff Induction is just Ockham's Razor for the Turing Age -- so there's no real challenge in coming up with an exoteric framing. Sequential Decision Theory can be framed quite simply as well: If you know the outcome of all choices available to you (provided by Solomonoff Induction), Decisions become trivial. The reason I'm hammering on this is that the failure to understand lossless compression's value as the intrinsic utility function of unsupervised learning has untold opportunity costs to society: The enormous resources poured, not only into the social sciences but social "experiments" conducted on vast populations without any serious notion of "informed consent", should be informed by the lossless compression of a wide range of longitudinal social data. Google DeepMind should be at the forefront of this given its background and Google's resources. See this question I put to Kaggle: www.kaggle.com/general/37155#post207935
Slides: ranzato.github.io/
Really useful! Thank you
Thank you so much for your interesting presentation!
Very recommendable! Talks by two very renowned researchers
Great talks but wish Alex Graves had paced his talk better to focus on the interesting stuff instead of the more well known ideas.
Not sure if you guys gives a damn but if you're bored like me during the covid times then you can watch pretty much all of the latest movies on InstaFlixxer. I've been watching with my girlfriend for the last couple of weeks xD
@Torin Carl definitely, been using InstaFlixxer for since december myself :)
Maybe it's possible to use narrow AI to figure the optimum method of unsupervised learning for us.
What is density modelling?
modelling -> finding out /predicting -> Basically finding out from data a model
density -> here it means probability density function -> i.e. probability distribution of data/thing to be modelled.
Amazing
He does not fully understand, I think.
thanks god
getting the metric right seems like feature engineering...
0.5 < P(the cat sat on the mat | google talk) < 1
About 17 minutes and I had to stop listening because I felt like I had lost about a standard deviation IQ. Hasn't this guy ever heard of Solomonoff induction? Hasn't he ever talked to Shane Legg? The intrinsic motivation is lossless compression and if the agent is active the decision theoretic utility determines the explore exploit tradeoff as in AIXI. If passive it just compresses whatever it's given as data.
Thats Alex Graves... well renowned at DeepMind. He's released papers with Shane Legg.
Fish !
@@theJACKATIC I listened to the rest and he did, finally, bring in compression as one would expect of someone with his background. And it does appear important. His presentation threw me off. At a meta level, he really should start with the "high level coding" of his presentation: Describe the space in terms of AIXI's unification of Solomonoff Induction and Sequential Decision Theory before breaking down into his 2x2 taxonomy. That way it would be clear that "unsupervised learning" is simply lossless compression toward Solomonoff Induction's use of the KC program's "latent representations". He appears to have his head so far into the techniques of lossless compression that he elides the "top down" definition of AGI as the start of his "high level".
@@jabowery some persons are not so advanced in this field and starting from the known to the unknown is a great technique in passing knowledge.
Great presentation🙋
@@coolmechelugwu7305 Solomonoff Induction is just Ockham's Razor for the Turing Age -- so there's no real challenge in coming up with an exoteric framing. Sequential Decision Theory can be framed quite simply as well:
If you know the outcome of all choices available to you (provided by Solomonoff Induction), Decisions become trivial.
The reason I'm hammering on this is that the failure to understand lossless compression's value as the intrinsic utility function of unsupervised learning has untold opportunity costs to society:
The enormous resources poured, not only into the social sciences but social "experiments" conducted on vast populations without any serious notion of "informed consent", should be informed by the lossless compression of a wide range of longitudinal social data. Google DeepMind should be at the forefront of this given its background and Google's resources.
See this question I put to Kaggle:
www.kaggle.com/general/37155#post207935