DeepMind - From Generative Models to Generative Agents - Koray Kavukcuoglu

[@RaviAnnaswamy]

The key seems to be that the setup provides a “grounded bottleneck for the learned representation” by requiring that the representation learnt is not a description but an action plan hence connection to reinforcement learning In one of the Turing lectures from 1970s Simon Newell I think distinguish between two ways of describing a circle A circle is a set of points equidistant from a point A circle is the path traced by a point moving around at a constant distance from a given center The first is a static description the second a procedural description While the first type can guarantee accuracy it cannot guarantee minimalism and hence generality If VAE and GAN are of type one impala spiral is type 2 So this approach looks at a picture and generates a program (plan of actions) to generate it; it sees a digit and infers strokes needed for it Since the real information content of an mnist digit picture is far less than 789 bits the representation learned by spiral seems to be far economical This has mind boggling implications Btw such brilliant delivery amazed by precision of thought and words throughout

[@RaviAnnaswamy]

Also listened to prof bengios presentation at Microsoft on disentangled representation learning and prof Lecun additional approaches to unsupervised learning examples..2018 seems to have begun a new direction and already delivered fruits in unsupervised learning and this marriage of unsupervised learning with reinforcement learning is awesome

[@dr.mikeybee]

Fascinating. Using various models as components in training seems like the best way forward. Brilliant.

[@citiblocsMaster]

Koray is an archon of Dan Ariely and George Clooney

[@berkk1993]

i thought it was George Clooney

[@EngIlya]

No links to the research paper he was telling about?

[@GuillermoValleCosmos]

What is the explanation of the contrastive loss in 16:30 ?

[@GuillermoValleCosmos]

Ah I see, there is a typo, the second term in the loss, should have two different conditionings c_1 and c_2. See here: arxiv.org/pdf/1711.10433.pdf#page=6

[@utomo8]

Can impala have Good Natural Language Understanding ? Natural Language Understanding have very big potential in the upcoming year

[@litman3980]

türk olarak gurur duydum

[@ilyadaaa9284]

🇹🇷

[@enesmahmutkulak]

Bu alanda çalışan Türk'lerin olması ne güzel.🇹🇷

[@oldschoolgreentube]

We are engineering our own extinction.

[@eyeofhorus1301]

+oldschoolgreentube newbz

[@vcool]

Luddite

[@oldschoolgreentube]

Absolutely. @@vcool

[@FacePalmProduxtnsFPP]

Gross

[@MONOLITH-yd4vq]

FacePalmProductions just think people want this 🤖

[@FacePalmProduxtnsFPP]

MONOLITH 2045 I'm sitting here scrolling through it again trying to remember what it even is...

[@FacePalmProduxtnsFPP]

MONOLITH 2045 now I remember why I said gross, I lost interest so fast.. 😂 Jabril and Siraj Raval are FAR more interesting and fun, and equally as informative.

[@shubhampateria2267]

FacePalmProductions People like Jabril and Siraj are not qualified enough to speak at ICLR!

[@FloydMaxwell]

Most of these DeepMind talks are like that joke about the Comedians convention. They were all so familiar with jokes, they had them numbered. Someone would get up and say "217". Much laughter. Someone else would get up and say "133". More laughter. Then someone got up and say "351". Hysterical laughter. A newcomer asked why the third person's joke was so funny. Answered another comedian, "It was a new joke." - - - - - The point is that unless you invest more time in using analogies, and front-loading your talks with the reasons to listen to it, you just end up talking code to people already familiar with the code.