Professor Yann LeCun visits MBZUAI

We had the pleasure of hosting the esteemed Turing Award-Winning Computer Scientist, Professor Yann LeCun.
During his insightful session titled ‘Objective-Driven AI: Towards Machines that Learn, Reason, and Plan', Professor LeCun engaged our faculty, researchers, and students underscoring that a major evolution in AI is still to come and its impact will improve collective human intelligence.
Watch the full talk.

[@r.s.e.9846]

Thanks for releasing this.

[@kyungyangpark8339]

Thank you for sharing this. I am on the same page regarding your Future Universal Virtual Assistants.

[@zawyehtet3716]

I have been enthusiastically applying the university's admission since 2022. She makes me a sense of future AI.

[@tarchearkhe3825]

So great Yann Lecun , in 2015 I Said that we should replicate the say we learn since baby age , exactly that / Same for Memory we need other architectures like global flash one , full memory path in one shot and only the one goes up

[@mnlnl65091]

Yes. Human brain is not a single model but a handful of specialized models commanded by a coordinator. It is not fair comparison between LLM corpus size with visual signals. A 4 years ago only knows about 500 words. I hope the next break through is persistent memory for knowledge given Transformer effectively introduced working memory in attention. Today LLM is passive, in a sense it is trained to answer questions. Human has self-supervised learning and reinforcement learning capabilities. For a break through such as creativity, train LLM to ask questions. All teachers know good students are who are curious.

[@tarchearkhe3825]

Yann you said many times: IMAGINE. This is the key : machines need to imagine , to dream, do you see what mean?

👍 Great !!!

[@benjamindilorenzo]

so, why is Yann still pursuing end-to-end gradient-based world models? Why not use other approaches that have been proven to be highly successful for reasoning and planing such as neuro-symbolic systems in which reasoning is done via Knowledge Graphs.

[@XX-vu5jo]

Its the current sota that is proven and trusted for now. There are others but are still considered a proposed approach or very theoretical. He is focusing more on what is trusted for adaptation.

[@sapienspace8814]

@ 1:13:20 I completely disagree with Yann on RL, whenever is a world model inexact? All the time! Everything in life is inexact, I do not know anything that is not inexact. Simulation is where most of the learning can train the weights and RL is made efficient by approximation (membership functions). Perhaps he has not read the new chapters from Sutton and Barto's book on Reinforcement Learning (RL) in their 2nd edition.

[@user-mo9fz4tk2r]

"modern democracy without books" -> Australia.

[@alsaderi]

My hope is in Yann👏🦾💙❤