Dileep George: Brain-Inspired AI | Lex Fridman Podcast

[@lexfridman]

I really enjoyed this conversation with Dileep. Here's the outline: 0:00 - Introduction 4:50 - Building a model of the brain 17:11 - Visual cortex 27:50 - Probabilistic graphical models 31:35 - Encoding information in the brain 36:56 - Recursive Cortical Network 51:09 - Solving CAPTCHAs algorithmically 1:06:48 - Hype around brain-inspired AI 1:18:21 - How does the brain learn? 1:21:32 - Perception and cognition 1:25:43 - Open problems in brain-inspired AI 1:30:33 - GPT-3 1:40:41 - Memory 1:45:08 - Neuralink 1:51:32 - Consciousness 1:57:59 - Book recommendations 2:06:49 - Meaning of life

[@eduardoluizrhein4983]

Hello Lex :) you could make a video with Alan Edelman or someone responsible for the development of the new JULIA programming language :) it would be really cool, since Julia is going to “compete” with python

[@pebre79]

Gold pure gold

[@sagarpatro2790]

Please bring Anil Seth and discuss conciousness.

[@DavidPeebles]

Hi Lex. I would love to hear you interview these AI people: Ken Forbus (users.cs.northwestern.edu/~forbus/), John Laird (laird.engin.umich.edu/), Ashok Goel (www.cc.gatech.edu/people/ashok-goel), and Pat Langley (cll.stanford.edu/~langley/).

[@boyzrulethawld1]

Keep the advice section going Lex! I realise u asked here, but pls keep asking haha

[@artfuldoc]

I have been a huge fan of Dileep's work ever since his time at Numenta. Thank you so much for this interview Lex. Would love group interviews. For example Manolis Kellis and Dileep George in one discussion. Putting people from different fields together and listening to their take on similar concepts can be intellectually satisfying.

[@TheStainlessFish]

Pinnalade!!

[@thisisajaym]

njan angotu parayuvan povarnu! apozhekm Suneej chetan angu vannu

[@hoka-v3h]

@@thisisajaym .,.,pl

[@hoka-v3h]

Mll

[@onerib781]

This was the most mind-blowing thing I’ve heard in a while. Seeing how the visual cortex encodes concepts into cortical columns, how those columns are made up of thousands of neurons, how the columns connect to other columns in a layer, the multiple layers, how all sensory information goes thru the thalamus, then distributes this information to other parts of the brain, how our brain is constantly creating models of the world and this model is constantly being updated thru feedback connections and sensory input and our expectations, I guess. Best podcast ever, for me

[@asdfkjhlk34]

It is amazing, great summary

[@boyzrulethawld1]

Funny how our brain is blown away by rough explanations of its own workings, how fascinating it must be to truly know ourselves ;)

[@timf9740]

Liking the 'book reccomendations' section Lex! Please keep it up going forward.

[@tyfoodsforthought]

Second this! Pls for the love of reading, keep up the book recommendations!!!

[@scgamer008]

Lex, thank you for the great content! It would be greatly helpful if you could put the recommnended books list into video info section. Thanks!

[@timf9740]

@@scgamer008 I 2nd that

[@chinmayjoshi4235]

Definitely.. allows us to explore the subject to a deeper level

[@MOSP14]

Where is it?

[@luketambakis4216]

Lex is on a RAMPAGE with these conversations

[@pebre79]

Thanks!

[@mosca-tse-tse]

Fascinating conversation!!!! The enthusiasm of both gentlemen is contiguous. This is the kind of research I would love to contribute to. Thanks so much for this episode. 👍🏼

[@foreverseethe]

Damned Google spell checker.

[@ashwinshankerg]

"Kuriakoseeeeeeeee" malayaleeeeee! wohoooooo

[@rohitvishnu666]

Malyaliyiii ✌️

[@dennis_johnson]

Represent

[@ashiktm4188]

😁

[@fazil47]

+1

[@pauljnellissery7096]

Malayali

[@torarinvik4920]

Lex is such a diverse scientist, he includes multiple different fields like psychology, physics, linguistics, computer science, biology ect in his view of holistic knowledge. Truely something to be saluted.

[@abc-co7fy]

അങ്ങനെ ഒരു മലയാളിക്ക് Lex Fridman-ൻറ പോട് കാസ്റ്റിൽ ഇടം ലഭിച്ചു. മലയാളികൾക്ക് അഭിമാനം നമ്മുടെ ഭാരതത്തിന്റെ സ്വാതന്ത്ര്യ ദിനത്തിൽ.

[@rejy92]

മലയാളികൾക്ക് ഇത് ആഘോഷരാവ്

[@darthmoomoo]

@@theyeking7023 It is.

[@aus5882]

പിന്നല്ലാതെ!!

[@thedeathstar420]

intense script

[@matts3414]

Microsoft translate says: "So a Malayali got a place in Lex Fridman's podcast. Proud of Tamil Nadu on India's Independence Day."

[@rejophilipjose7763]

Just love when Lex just asks all the questions that you wanted to be asked 😃.

[@tgwashdc]

Dileep displays great insights with unmatched humility. Comparing CNN and brain's "phobia" to underweight peripheral vision knocked me off (time line 1:15). Spasiba Lex, for yet another great interview.

[@crab-dogjones4659]

It occurs to me that motivation is the ultimate driver of human consciousness. I'm not sure you could build an authentic human AI without making it hungry and afraid.

[@NS-gr9cy]

great point and also put in rest of the emotional mess and you have a human. No one would suspect.

[@PrashantMaurice]

As Naval puts it, every desire you have is an axis where you will suffer. In these network's case its desire to classify cats vs dogs.

[@Naentrikakudapikalev]

@Pedro Abreu Saw matrix movie ?

[@monkeybusinessmediallc9373]

Also, all animal brains evolve in response to environments. What is the environment that shapes the evolution of A.I.?

[@foreverseethe]

@@monkeybusinessmediallc9373 greed, fear, curiosity and competiton. Bad news.

[@tyfoodsforthought]

This is exactly my cup of tea. SUPER excited to watch! Thank you, Lex!

[@alwadud0196]

Awesome, but how is that cup of tea connected to you?

[@torarinvik4920]

What I came here originally to post/ask. Why could I not find a wikipedia page of Lex? If it does not exist, can anyone please make it? He is so worthy of this and more!

[@vaibhavsingh-ro5tl]

Lex tell about your startup man!

[@wuwonder2797]

Great video! learn a lot from the interview.

[@DominicDSouza]

Wow! Every time I watch these podcast I am blown away. Thank you again Lex for fabulous content and topics. Also, thoroughly enjoyed hearing from Dileep and his approach.

[@TheSocialDrone]

Very interesting conversation, complex concepts explained in simple language!

[@ScienceAppliedForGood]

It was a very insightful talk by Dileep. Thank you Lex for having an interview with him.

[@vast634]

The most interesting AI related interview to date.

[@vanpavlic159]

Thank you for all the amazing podcasts. I'm excited for the future. You can only get better at this, and you are already the best. Keep asking great questions.

[@edenaut]

25:00 The gorilla experiment shows the feedback mechanism very good I think. I think 'seeing' is very goal-related (conscious or unconscious goal). What somebody value more is what he sees more, OR what obstacle comes in the way of that goal

[@alexfedorov5774]

The guest is a good one. His understanding has analogues in (my own understanding of) machine learning. One of the models of the brain that George mentioned, "brain has multiple hypotheses that compete with each other, and the one that wins explains itself to others", is like expert distillation, where "expert" is "repeat this computation N times and pick the best", and "distillation" is "minimize the loss between what we said and what the expert says". Modern AI can do that, with sequence-to-sequence models such as the Transformer acting on observations to produce behavior (program), and there is no need for symbolic hypotheses. Imitating the best version of yourself is no good when no version of yourself is good, though, so it may need pre-training. The feedback connections (that are more frequent than feedforward connections) is like the loss minimization, the basic primitive of machine learning. In standard formulation of machine learning, having so many feedback connections is like summing over arbitrarily many individual prediction errors, and backpropagating that (can be re-formulated as a "add a misprediction loss here" function). Modern AI can do that. 39:20 Top-down controllability sounds like routing differentiable connections through a program, and have decisions be made based on that differentiable context (then have the misprediction loss be computed per-decision, meaning that each "down" low-level decision predicts some decided-at-the-"top" quantity and maximizes it, like in reinforcement learning). Neural-network-based operations can be embedded in arbitrary programs as long as their output computes and backpropagates some loss, so, modern AI can do that, in the sense of "it can theoretically be programmed". And to add a more philosophical view on this, if personality is what determines decisions, then personality is just a neural representation learned to maximize important quantities. And to add more, such a learnable scaffolding could be a better replacement to random-search methods like evolution. I mean, would you trust your baby to randomly-generated numbers signifying behavior, or to thoughtfully-picked numbers? Learning on-the-fly quickly enough is a real challenge, though. But maybe keeping neural networks very small and relying on all those other things instead of neural network size is the answer here. However, the "all" in that sentence is like all programs, which can't just be put in one program with one big switch, and has to be generated. Generate everything, and everything should fall into place. Generation of all programs, with the things above included: it seems like that could be AGI. 1:53:30 Self-awareness could be said to be a neural representation learned to maximize important quantities. Those important quantities, goals such as urgency, can be modified based on something already learned, such as fear of death. Self-modification should be included in generating everything, meaning at least a "thoughtful" decision to re-generate, to know when to deviate. My nose is bleeding, though. Madness. Hope you're having a good time.

[@bloodcarnage8285]

I am studying data science and this conversation helped.

[@codyboe4737]

I think more people need a reality inspired life. Love the podcast, amazing concepts.

[@NeuroReview]

Rating: 7.8 In Short: The Mind is Powerful Notes: This was a really neat neuroscience/AI combined discussion from someone who is thinking about how to design the mind in AI and programs. Dileep is a good speaker and interesting guy, and explains the brain at a really nice level for those that might not be familiar. He does it with an almost childlike enthusiasm that is super refreshing in the realm of science. They start with some nice basic science that reminds me of Cricks ‘Astonishing Hypothesis’, basically through explaining the visual system you learn about the miracles and beaty of the mind. A cool idea he explains is that in order to make something like the brain, you need a well made perceptual system. Basically percepts are necessary for concepts. This is a cool fundamental idea that argues that further understanding of our senses, i.e. vision, can help us understand deeper aspects of the brain. The end was kind of cute with books and meaning, very short, seems like Dileep is a classic nerd/researcher. A great convo for those interested in Neuroscience.AI interface

[@wentianzhao64]

This is such an amazing talk! Cannot agree more on the point that perception has to be integrated with cognition system before truly solving the perception problem.

[@mtumasz]

Great mind & ideas. As a person trained in psycholinguistics on Chomsky’s universal grammar I thank you for offering a different perspective 🙏🏻 However I seriously doubt existence of different memory systems in the brain. This 20. century computer metaphor must end! There’s one memory system processing everything in parallel and continuously playing with properties like recent, important-for-survival, repetitive, etc! And it’s all inseparable from areas like emotions, vision, hearing, spatial awareness...

[@B33t_R007]

6:10 "initially they where just trying to simulate a cat brain" - "to understand the nature of evil" hahahahaha. everybody in the office was looking at me

[@pebre79]

Yes been waiting for this interview!!

[@davidfell5496]

Having been exploring the idea of AI actors that could lie, and thinking about approaches to implementation, there are many, many fascinating thoughts from Dileep here. Very very interesting. Thank you.

[@JohnTigueDotCom]

Since you interviewed Hawkins and George, might as well go upstream to Pentti Kanerva, author of Sparse Distributed Memory (1988). Start with his 2009 field intro/review 21 page paper: Hyperdimensional Computing: An Introduction to Computing in Distributed Representation with High-Dimensional Random Vectors, (Cognitive Computing, 2009 1:139-159), www.rctn.org/vs265/kanerva09-hyperdimensional.pdf, DOI 10.1007/s12559-009-9009-8

[@trusteddie7]

Great podcast!

[@LaEl1010]

I don't even have to watch this to know Lex only puts out FIRE! Listening in 3...2....

[@quaidcarlobulloch9300]

Wonderful topic and interviewee Lex! Thanks for the great content.

[@janbosenberg107]

"This was so much more fun than I expected" I think you speak for us all. What a wonderful interview.

[@jasonsebring3983]

I always felt there was a bit of woo-woo in "emergence of consciousness" and Dileep is taking a pragmatic approach to avoid that. Right now it feels like we are looking at a result and anthropomorphizing it to meet our hopes. Yah it does fool us but it doesn't understand anything still.

[@funnyvideos-funnyoutloud2618]

Happy Birthday Lex!

[@kimchi_taco]

1:26:00 Perception is the first thing you have to build but the last thing you will actually solve.

[@ivanbiji]

Proud to be Malayali

[@Goat-e3g]

Malayalikale nanamkeduthaelle

[@deeplypresent]

Fantastic conversation Lex and Dileep! I didn’t realize how close we are getting to having a conceptual model of the brain. It will be very interesting to see a model built that connects all of these ideas together. Although not stated specifically, I heard many references to the “free-energy” principle, one of my favorite explanations of how the brain does probabilistic error correction between our internal models and the external world. I also think that an important component to intelligence is emotion, otherwise you get a Buridan’s donkey situation. One part where I disagree, I believe humans do have an innate set of goals, which we share with all life: to survive and replicate (the fundamental goal being to propagate our genes). The amazing thing about humans though is our desire and capacity to understand the world and it’s amazing to see you two working so diligently to expand our understanding in these areas. Definitely earned a subscription from me! Keep it up!

[@ilikenicethings]

How is it possible for there to be a thumbs down on this video which was released 1 hour ago but the video is 2 hours long?

[@bigbabarian3667]

Not finished but still mind boggling ffs! Excellent channel 👏 👍 love it 😀 😘

[@bigbabarian3667]

Was the JRE that brought me here btw 🍺

[@robinampipparampil]

This is brilliant! Thank you very much Dileep George and Lex Fridman

[@ohannesemirzeian2024]

HAPPY BIRTHDAY! GOOD LUCK!

[@joshGFLS]

Don't forget to support Lex on Patreon. $3.14 a month... imagine if Lex could fly guests out or go to labs.

[@adi-kr6sh]

lex podcast are a gift for humanking Evolution and education

[@travisfitzwater8093]

Just follow every moment of a child's development from birth to 18. Track the images the child sees and the words it hears. Track everything the child says along the way to correlate its learning from it's experiences. This will show how to train a neural network about things whether they are formally logged or written down somewhere informally.

[@angeltafolla9783]

Awesome Video!

[@quasarsupernova9643]

Wonderful guest.

[@satychary]

Pretty cool guy, lots of good insights.

[@exacognitionai]

Excellent content but barely scratched the surface. Wish the questions had dug much deeper into the details & bleeding edge of cognitive AGI. There is so much fascinating work going on in the space.

[@empathylessons2267]

Oh hey cool, my literal career goal. This is gonna be exciting!

[@abhilashasaroj204]

Great conversation !

[@jamesr2936]

Love Lex's podcasts. Is it just me or whenever he asks that final 'meaning of life' question he stumbles all over it!? Haha

[@gnarlyandy1]

Does he stumble over it on purpose, So to imply he does not know the meaning of life.

[@kumarvikas_134]

Why did you edit @ 9:32 ?

[@franktfrisby]

Dude I love these Podcast. Great Work!

[@igorfelipezampieri7137]

Brazil here!!!🇧🇷🇧🇷🇧🇷

[@ハェフィシェフ]

Does anyone have the timestamp for amortised inference? Edit:found it at 58:13 if someone else quickly want to find out what it means

[@agiisahebbnnwithnoobjectiv228]

The most important question in the end. Objective function of the brain/life. Knowing this, its easy to solve AGI. Watch my 2 part video for answer.

[@NielLai86]

I think Numeta is hitting onto something good. I got gut feeling they will have a break through soon. Covnet and backprop is not gonna cut it.

[@vast634]

Their approach looks way more promising than other neural network architectures I have seen. Especially if it can learn without humongous amounts of training data provided (wich would hit a hard ceiling in areas where those data collections are not available). E.g. brute force learning...

[@ericshetka4898]

I was having a difficult time understanding him at first, but got the hang of it after a while. I think if this guy were able to separate tonal human communication and vocabular communication while in a state of derp breathing, and went back and recited the English nursury rymnes and some poetry in English while deep-breathing, he has to potential to "snap" into speaking perfect English with his choice of accent within a year.

[@drew4054]

Your the man Lex

[@HeliTom84]

Interesting👍🏼😊 good job

[@SS-qf3pq]

If the motor and visual cortex is an important prerequisite for the language cortex, how do people with disability at birth (mainly visual) develop their language? Any thoughts on that?

[@NishanthSalahudeen]

Awesome stuff

[@Jasonchaaang]

Yoo lex! How come no videos with anyone from Mobileye? It seems at this point mobileye may be in the lead with full autonomy

[@NishanthSalahudeen]

My question is... will an agi comeup with or accept the concept of god /religion? Will be weird if a computer starts a religion and recruit humans. The reason for faith seems to be built into the way we process information... "causality and agent seeking". So if we model agi after humans, this bug(or feature!) would follow as well

[@billcowhig5739]

Are transcripts available?

[@citizizen]

Hi, i was thinking about neural circuitry and the following occurred: you have single OR's, and this might be a representation of something complex. More then one OR for an object, is not something feasible, simply gibberish. Now, these things that have a single OR to it might be complex and deep phenomena, i call it a picture (singular picture). Those are totally explained and thus possible to be absorbed if such a picture is a single phenomenon (a single OR). If there are faults, then one does not have a single picture. If there are more then one 'or' to a picture, it simply makes no sense. However, i guess we do our circuitry on this level. NOTE: if one learns by failure, to fail with something that is learned by failure might mean, better connections, perhaps reasoning even (multiple OR case).

[@pebre79]

Goals are just justifications of a humans primal directives: fighting, fleeing, fornication, hunger avoidance, social acceptance, etc.

[@simonbean3774]

Y'all playing with fire

[@arunavaghatak8614]

Please elaborate.

[@simonbean3774]

@@arunavaghatak8614 This shit will replace humans. AI prometheans

[@meetboy352]

In which chapter of the podcast does Dileep speak about the spread of the COVID-19 virus?

[@IlyaRoy]

This sounds awfully like Ray Kurzweil's model described in How To Create a Mind - 2012.

[@kevinhero66h95]

How, I wonder, do the concepts of visual learning and conceptual modeling apply to the neural function of a blind subject? The idea of how these neural systems interact in an environment of sensory deprivation. Does isolation of sensory stimuli give us more insight into the nature of raw cognitive function? I am sure papers have been done on the subject, but I am not involved in the academic community.

[@BHBalast]

I'll try Bubble, funny because I already understand a big part of this poem, cause I'm from Poland. Also I already know that second part of Russian's language. :)

[@eduardoluizrhein4983]

Hello Lex :) you could make a video with Alan Edelman or someone responsible for the development of the new JULIA programming language :) it would be really cool, since Julia is going to “compete” with python

[@Kaget0ra]

"You're assuming I'm human," hahaha.

[@aidanthompson5053]

Vicarious Probabilistic reasoning for intelligent systems The minds eye

[@SahilDawka]

Does anyone know of any research in the vein of bringing up a brain in a controlled visual environment (say, without straight edges) and observing differences in cortical structure?

[@thisisajaym]

👇🏽 Malayalees attendance button

[@raresmircea]

Please invite Andres Gomez Emilsson from Qualia Research Institute 🤘

[@Skindoggiedog]

Lex. Love the podcasts. Learn how to say 'understand.' It's not 'undershhhstand.' Thanks.

[@SamuelJFord]

This was one of the best!

[@jeffin8029]

Lex love asking the meaning of life

[@tommole645]

Class 👌

[@nodesofnature603]

46:00 I wonder if AGI would laugh the same way.

[@BILLY-px3hw]

10 seconds in and they are already ripping on the blue brain project..."you can't make a brain without understanding how it works" Me: well that didn't take long

[@werner_s]

if the blue brain project is successful they get an artificial brain they do not understand.

[@davids9522]

He said what color is the letter e in Google on the google homepage. I can recall every color of every letter correctly. Whats that say about my brain? Also Id like to add, I can remember things from 1 years of age. Weird enough my first few years of memories are in black and white. Also, I remember when I was a young child I could look at someones face, close my eyes and still see their face in perfect detail.

[@monkeybusinessmediallc9373]

Just wondering, what's the difference between neural ion gates and ion channels, and Maxwell's Demon?

[@neelpatel3844]

Is Aquafina is sponsor?

[@jaredbeckwith]

Love brains and artificial brains! ❤️

[@neelpatel3844]

Dileep ‘doctors wear underwear’ George

[@adilzamal3218]

I miss his window

[@azsegrxdhtfgvijnkomlewrhtg9508]

41:17 I knew that one.

[@Loom-works]

Facebook is the simulation of mind . The person who runs the A.I. Facebook program is the same person who created the SIMS Game. The super computers are creating a database of thought process. Data base of thought process within individuals and groups. This has gone as far as a physics game. Equations of interactions , create GATEs that can utilize individual mind process by implanting information . This is called ( The Game) in the CIA program running. I will add the information to some of this information that you might like.

[@Loom-works]

kzbin.info/www/bejne/aXLOZohoh66cebM

[@purplemashine9122]

Here's the problem: The exponential rate at which AGI,overall tech in general is rising faster than we know about the human brain. I have a solution: Put. Together the latest research on the brain ,plus latest normorphic AI,photonic AI, all the deep ,latest deep reinforcement learning,the billion neurons on a AI on a computer Source: Itbreifcase.net(as of last week) all the other breakthroughs ,and see what happens.

[@purplemashine9122]

@@ahmadayazamin3313 I hear ya. As far as systems being trained on less data,generalized scenarios it has NOT been trained on? They already EXIST IN labs all over the globe.Thts EXACTLY the point of my research: Google ,others have been releasing info on this since jan 2019. It's frustrating to explain to the world the AGI ,singularity has begun ,yet no-one listens ,bothers to even investigate it further,ask to see the whole research,when I PROMISE u,when u see all the research,it WILL INSTANTLY change your life in a PROFOUND WAY! I call it the singularity Puzzle - pattern,it includes EVERY AGI breakthrough from jan 2019 until now. Every mind blowing robotics ,brain computing interface ,VR ,quantum computing ,over all deep science - MIND bending breakthroughs. Again,when u see it all together ,it will be REALITY SHIFTING! (No articles are hype, fear mongering perceived ,or projected views, I ONLY COVER REAL ACTUAL Science breakthroughs in labs all over the globe)

[@ahmadayazamin3313]

@@purplemashine9122 I see your point. After all, who would have thought that a computer could play the board game Go at superhuman level a decade before the speculated time. Still, I would say the time it takes to actually reach this level of intelligence would be at least a decade, since the mainstream research labs at big tech companies don't seem to get the point that smaller teams / startups are trying to make through their breakthroughs. Yes, the technology exists (of which I would include the Recursive Cortical Network mentioned in this episode) and we would be bound to achieve these heights.

[@purplemashine9122]

@@ahmadayazamin3313 Did u know tht Google created AI that EVOLVES on ITS OWN? Did u know that a robot hand has conducted 600 scientific experiments....ON IT'S OWN ,(All In labs) ? I mean it's just 2 of MANY of such articles I have gathered. At what point can we say,yes we've "arrived"? Again based on my research ,we have already. Hope one day to share with u ALL the research,again it will Change your life for sure.

[@tbainbridge]

The nervous system came about by evolution. It is the cumulative outcome of millions of years of trial and error. A microprocessor is in no way the same level of complexity as the brain. A microprocessor is simply a very long way to do problems running at very high speeds. The brain is an elegant design in its architecture itself. It is so much better wired and handles high dimensional data. So the brain doesn't need the super high clock speed a computer does. Because a computer, at its base, uses very simple logic gates and machine code. It needs to go very fast because it is always working in a simple serial way. The brains superiority is not in its software, its in its structure. You can see this even by understanding 2 point discrimination of touch. The minimum distance you can discriminate between 2 points is because of the structure of the neurons. There peripheral neurons will send feedback in different coding frequency to the central neurons, the difference in signals tell the brain how to orientate where the points are. That structure is the intelligence. And the only way to get there is evolution.

[@bcisec2184]

Wouldn't we be able to create some sort of simulation of the evolution process using machine learning? Isn't evolution (in theory at least) a sort of unsupervised learning process happening at a biochemical level?

[@MaxAravind]

Angane ivideyum malayali vannu

[@pauljnellissery7096]

Sathyam