Ian Goodfellow: Generative Adversarial Networks (GANs) | Lex Fridman Podcast #19

[@lexfridman]

This conversation with Ian led me to rethink the way I see several basic ideas in deep learning, including generative models, adversarial learning, and reasoning. I definitely enjoyed it and hope you do as well. 0:00 - Introduction 5:40 - On cognition/consciousness 7:04 - Self-awareness 14:00 - Deep learning 23:10 - Limits of machine learning 26:44 - GANs 42:32 - Semi-supervised learning 49:34 - Data augmentation

[@lexfridman]

@Modestas Jakuska Soon! We had a great technical conversation about theoretical computer science.

[@lilibalister]

Cool insights on GANs :D Happy idea : Interviews with artists/designers working with GANs maybe Mario Klingemann, or others there's been like a boom of AI Art recently since Ars Electronica AI in 2017

[@danielbright3340]

goof stuff Lrx ,Thanks

[@yehchihyu9970]

@@lexfridman Dear Lex Fridman, would you have an interview with Chris Lattner?

[@PATTHECATMCD]

A little dissapointed that an AI scientist bemoans the need for big data but doesn't seem to mind how skewed or biased that data is. Compare to a human listening to a politician. Sooner or later the human realizes the politician is just feeding them falsehoods to retain control.

[@Jone952]

You should've animated Ian's photo with a GAN after the battery ran out xD.

[@xoxo-sf1zg]

:o

[@gomes8335]

Oooo

[@Jone952]

@@AirSandFire thanks I try my best

[@MuhsinFatih]

maybe the reality is that they were doing that, and had to publish right away so didn't have the time to generate the rest of it. (queue "alien conspiracy.mp3"...)

[@armin_hammer_studios]

I came to the comments to make that same point - it’s nice when separate models converge :)

[@SY-me5rk]

Fascinated by how well Ian communicates ideas, almost too well.

[@xoxo-sf1zg]

@@iphgfqweio well yes, but actually no.

🤔

[@D9Wx]

@@Neptunion118 like elon musk 🤔

[@Z0Gab]

Its a symtom of being smart

[@BenGras]

i agree. i was just thinking how impressive it is that he never says 'um.'

[@Kakerate2]

5:40 On cognition/consciousness 7:04 self-awareness 14:00 how do you begin to summarize Deep Learning? 17:09 what is Deep Learning? 18:28 Example of Deep Learning that is not Gradient Decent on Differentiable functions 23:10 Shortcoming of ML 24:18 Facts and Expert Analysis 26:44 origin of the idea of GANs 39:00 GANs efficacy 39:53 Brief history of GANs 42:32 semi-supervised learning 44:30 circumventing data labels 47:22 domain-adaptation algs 49:34 "Do you think GANs could be useful for data-augmentation?" referencing

[@reasonerenlightened2456]

deep learning is just a form of declarative programming.

[@whiteF0x9091]

This guy is extremely smart, knowledgeable and passionate about his field of study. A great interview. Thank you Lex for these podcasts !

[@Aureole62]

He's basically a founder of a subfield of Machine Learning.

[@dolphinwhale6210]

I was seriously waiting for the interview with Ian Goodfellow

[@JousefM]

Lex is AGI and knew what you wanted ( ͡° ͜ʖ ͡°)

[@purelogic4533]

I have had the opportunity to learn directly from Ian Goodfellow in an online session. He was so unassuming and objective and answered questions regardless of whether they sounded silly or not. Very impressed with how humble he is although he knows so much.

[@Hexanitrobenzene]

He is really humble: 42:50 "...I am co-author of this paper but i cannot claim credit for this particular part...", 47:25 "...the authors are really nice in citing GAN paper, but I know they came up with the idea earlier.." . As I joke, he is of the kind of people who should be cloned :) - smart, eloquent, humble.

[@jimlim9224]

Wow as a generative model researcher myself, I am super jealous!

[@santiagoespada7533]

Congrats Lex, the podcast is getting better and better

[@gomes8335]

Yacc

[@ekbastu]

Lex’s gradient descent is working well ;) It’ll converge at Joe Rogan.

[@sophiavirdi6909]

I watched this interview a few weeks back and I was completely lost and didn't understand what was being talked about. I've since read some of Ian Goodfellow's book Deep Learning and I've also watched videos and read articles that provide an introduction to AI, Deep learning and machine learning. Watching this interview again, I can piece things together and I'm not completely lost anymore. I have a long way to go but I'm happy I'm getting there :D

[@MaggotDiggo1]

I'm currently at the pretty lost part. Which one of his books do you recommend? The book named Deep Learning seems to be his most popular.

[@ThomSonnyYeah]

Working backwards toward understanding something is fun sometimes :)

[@reasonerenlightened2456]

deep learning is just a form of declarative programming.

[@ahujaavi13]

Of all your podcasts, this one has the best questionnaire.

[@Dragonblood94]

Finally a talk about the real tech of AI a little bit beyond the usual. Thanks I learnt alot.

[@sumantadas6532]

Two legends from MIT & Stanford University 😍

[@MrJohnaiton]

thanks a lot for this interview, both the questions themselves and Ian's answers make the information incredibly accessible. Even though I don't know much about the topic, I only had to stop to google a couple of times to take away a lot of stuff , a much more relaxing format to take in information

[@jstov]

Hey Lex, just want to say... I love your videos. Still need to watch most of them... but I love the ones I've watched. Please keep making them.

[@JousefM]

I love Ian! Your podcasts are getting better and better Lex, keep going!

[@edouard.mercier]

Congratulations for this interview, which I found brilliant thanks to the capacity of the Ian Goodfellow to share high level insights from his expertise, and thanks to the very qualitative and well designed questions of Lex Fridman.

[@lightb00k34]

Your podcast is my favorite podcast. Keep making them and thank you for making them.

[@eliastouil7686]

If you don't know what a GAN is it's explained at 30:30

[@MiloLabradoodle]

Great interview! Ian does a great job methodically illustrating some of the interesting GANS topics.

[@r.d.machinery3749]

I really liked Ian's summary of the problems with neural networks arising out of their reliance on labelled data.

[@TheAIEpiphany]

We only had the audio modality when you asked Ian to prove that he's not a robot. You gotta repeat this experiment in a true Total Turing Test fashion!

[@arjunraghunandanan]

I am currently reading his book "Deep Learning" to study ML and DL. Then I noticed that there this old LFP episode on it. Enjoyed seeing how much improvements have been occuring in the world since then. This was a very insightful session regardless of it being 5 years old.

[@gulllars4620]

Around 24:00, I think i recall the researchers describing one of the properties of GPT-3 over earlier and smaller models is it becoming much better at few-shot learning, which seems to me to be a similar property to what Ian was talking about here with working memory. I also think i recall another episode of the podcast where it was concluded that working memory in this way becomes an emergent property of large enough reinforcement learning models, as it becomes a more effective way to solve broader problems than not having it.

[@perfumedsea]

I like this interview. I was confused when I first read about CNN layers representing different messages like contour, angle, color whatever. But at that time I thought it was me that's too stupid but then when goodfellow, the guru expressed similar idea, it give me great relief :) Thanks for the series.

[@thesystemera]

Absolutely loving these talks!

[@DieMasterMonkey]

Another excellent interview - except when, at 52:26, IanBot's servo power supply runs out. Fortunately he was able to complete the interview via Bluetooth, demonstrating the resilience of his neural network.

[@kaiwang2924]

I learn so much from this conversation.

[@DayB89]

That was a great interview Lex. Congrats! I'm personally impressed that Ian could implement a working GAN in just one hour :O

[@Martinsp16]

That’s awesome thank you for sharing your work

[@joseortiz_io]

Dude! You are the man! ❤❤❤ You bring in these legendary experts of AI . great job! ✊

[@danielnofal]

Amazing to listen to this in 2024 and see that even them didn’t see it coming. Min 24 Lex says “chatting with a neural net” and laughs. We are doing it everyday.

[@casperhansen3012]

The part about being able to give feedback to a network is interesting. If one could specify beforehand, that a person has a high probability of having the same color for each eye, then that would help the networks lots!

[@danielpgp941012]

Grabbed so many things from this talk. Thanks so much. Will soon be using semi-supervised learning on one of my projects !!

[@dynamitesteel]

Lex you are such a good interviewer/podcaster. Thanks for all the amazing content on your youtube.

[@philoneill9865]

Very interesting talk. This is a fantastic series. Well done by Dr. Fridman. Just a tip: I recommend Panasonic E-Volta for a reliable, long-lasting energy source. There is nothing that a GAN hates more than a power failure.

[@sarsajjadiyan]

Where can I find a transcription for this podcast? Thanks for sharing this great work.

[@chocol8milkman750]

you can download the video and then run a speech-to-text NLP on it... (JK) I'd really like a transcript as well.

[@VitoSuave]

Amazing interview! Definitely need more podcasts like this in the world

[@spinLOL533]

I think Ian Goodfellow is a good man

[@chrisray1567]

Raphael He’s a jolly good fellow

[@KravMagoo]

Dude NEVER blinks.

[@jayjayDrm]

i'm glad that i found one blink at 4:19. i got really scared...

[@KravMagoo]

@@jayjayDrm That was a glitch in the matrix.

[@nadialexa]

@@jayjayDrm that wasnt even a blink

[@hamudi7779]

Thanks for your work. A interview with yuval noah harari would be interisting.

[@chocol8milkman750]

i second that! But does Yuval do AI research? I thought he was more of a historian.

[@hamudi7779]

@@chocol8milkman750At first im sorry for my bad English. Yes he is a Historian. But it would still be Interesting to see the viewpoint of a Historian who is familiar with the AI research. It would make a good Interview.

[@EdSurridge]

Yes to that I out of industry perspective asking Yuval to address criticisms of his thinking which could be sourced with a introduction soliciting them? Another i Expert of interest could look at AI/AGI military game playing and how apposing war AI's could overt war if given a diplomatic service bearing in mind games played results. ?

[@supersnowva6717]

@Lex thanks for the podcast lex! Do you know roughly when the 4th edition of the Artificial Intelligence: A modern approach book come out? I can’t wait to purchase it!

[@BeGunNer]

Really well done! Amazing interview!

[@metalim]

From 52:25 to the end of the video we hear Ian's voice synthesized by GAN.

[@ridhakiller]

This interview is so insightful

[@Travthewhite]

Thank you for the work you do Lex keep it up!

[@sanjiblamichhane]

Please make your podcasts available on Apple Podcasts. That’d be helpful. Thanks in advance.

[@zhen3356]

Around the 5 minute mark, Ian says that our earlier assumptions do not hold. It is not true that the earlier layers learn more simple representations like edges and the end layers learn more abstract representations aka grandmother cells. Please tell me if i interpreted this correctly?

[@VaShthestampede2]

That is what I heard him say too. Perhaps he was saying that such representations are not so tightly constrained to particular layers. Kind of like there isn't one particular piece of your brain that someone can go in and pop out and make you forget what your parents look like. There's a region where its most likely that doing a lot of damage there would have the effect, but the exact location is not well defined. At least that's my best interpretation of his words.

[@ricksminecraft]

Excellent talk - what actually is a grandmother cell? What would a grandfather cell be? Thanks!

[@deekshantwadhwa]

1:04:20 it has become a reality, OpenAI has accomplished it with their latest GPT-3 model, under the CODEX project.

[@WallarooonCaffeine]

Lex you absolute legend

[@derasor]

Ian Goodfellow is absolutely brilliant. He should have unlimited resources at his disposal.

[@mingamanga]

This guy is a legend

[@gavinlin6636]

Is the Goodfellow in the video generated by GANs?

[@chrismorris5241]

I really wish I had friends like this. Peoples eyes glaze over at 1/5th of this.

[@blackrockcity]

You just live in the wrong place. I know exactly what you mean. The Bay Area or perhaps Cambridge, Mass are good spots for you to check out. Maybe Hawthorne CA is becoming similar.

[@zhangrong1822]

Thanks for sharing this great work.

[@QifangZhao]

33:25~36.57: Ian talks about why generative models even works in depth! His insight about usefulness of CNN architecture and its implication that generative models are hard to work well in domains other than CV and speech -> REALLY interesting! Say, if some aliens/AGIs with much better hardware than humans' brain, then they can evolve in really amazing speed!

[@achunaryan3418]

Bro is levelling up with the camera freeze.

[@AdrenalineVideos1337]

This was probably the most interesting episode on AI for me as a complete novice in the field

[@achunaryan3418]

He is the most on point speaker in the show yet, after chollet.

[@kozmizm]

I love what he says at 50:13 especially in respect to medical privacy at 51:24 , and I especially love his ideas for eliminating bias at 52:12

[@NomadicBrian]

I like the summary approach to technical book writing is. The issue with references is that details change frequently. I also like to apply quickly to projects that I have planned so I can test what I learned as I go.

[@ravitsharma]

8:20 Foreshadowing of GATO?

[@simonstrandgaard5503]

Great interview. Where does one find detailed info about the C 4 10 problem?

[@BiancaAguglia]

I might be wrong but I think Ian's referring to the CIFAR 10 problem (or data) not the C 4 10. 😊

[@toufisaliba2806]

Hey Ian at 43.55, you said "to get to less than 1% accuracy" I think you either meant "over 99% accuracy" or you meant "less than 1% error rate" correct?

[@CharlesVanNoland]

I caught that too, and a few sentences later he did say "less than (etc) error rate" instead. I am sure that he recognized that he misspoke seconds after the fact but by that point he was already preoccupied with continuing his train of thought. Surely by that point he felt it wasn't even worth making an issue out of. We know what he meant ;)

[@savvys3782]

I was wondering about that too

[@RedPlayerOne]

great conversation! thank you!!

[@poppe-q4f]

Is it me or Ian never blink...

[@russm8193]

"Thinking that I'm a robot"... Love Garth. Wayne's world, party time!

[@Arjun-jt7yb]

Your podcast is awesome.

[@emrek1]

17:19 sanat nedir? var mı referansı yakalayacak :)

[@bagushardiansyah7967]

thank's for this podcast course... what i got is back to the basic neural network...

[@InfiniteCyclus]

Consciousness is the romantic, or ethereal view of perception. All you need to be conscious is to have perception. They are simply the same thing..

[@wtfhowbizarre1946]

How AI perceive human to human violence? violent crime is culturally unacceptable in most cultures, unless you're at war or it's for entertainment. And how do AI perceive their place in our human world? How would an AI perceive morality? If an AI robot could save a human life or save a fellow AI robot, which would it choose?

[@notrito]

This guy doesn't blink 😮

[@sprink88]

6:53 Lex asks if "consciousness and cognition can emerge from current types of architectures". Goodfellow's answer is audacious and fascinating in its immediate implications: "If you think of consciousness in terms of self-awareness, the fact the agent itself exists in the world, ... more limited versions of consciousness are already something we get with RL algorithms if they're trained well." Of course, as abstractions, we can think of agents as 'acting as if-', or displaying 'versions of-' all sorts of 'conscious-like' behaviors. It's built deeply into our way of describing the world to want to do so - we anthropomorphise everything, all the while understanding that when we say 'the cute robot fell over and hurt himself' we know that no 'one' was hurt. [cf zombie world] However, if we see a bat, for example, hurt itself or be hurt for whatever reason, we know there is 'something that it is like to be a bat'. Not so with the cute robot. But Goodfellow seems to be suggesting that, in an extraordinarily primitive and limited way (as he says), there is something ' that it's like' to be an RL algorithm. Fascinating. Interesting that Goodfellow prioritises compute and data as the vectors for 'scaling' this 'limited' versions of consciousness he posits. A question for Lex: You've have had Josh Tenenbaum as a guest lecturer for MIT 6.S099: Artificial General Intelligence [1], and you've talked with Stuart Russel on this channel [2]. Russel has argued that "ML is a fabric, and what's needed is architecture". Tenenbaum's [3] work with architectures like Newells SOAR [4], although has ML components is not ML in itself. The late and greatly missed Patrick H Winston's[5] 'Strong Story' hypothesis the 'gold star idea' behind Genesis system,[6] which uses 'a small collection of commonsense rules and reflection patterns' to 'demonstrate several story-understanding capabilities'. Karpathy's famous 2012 blog post "The State of Computer Vision and AI: we are really, really far away." is always a great lens to view AI progress... So... My question is: What breakthroughs have we made since that post in 2012, to bring us closer to an artefact that can understand all the nuance that is in that photograph, and to be able to reason about it, and tell stories about it? How do you see the importance of DeepNN in all their forms along with compute++ and data++... Vs... architectures like those mention above, and what might be termed GOFAI? Thank you for sharing your work with the world. [1] kzbin.info/www/bejne/bYOylp-Pq9RrnJI [2] kzbin.info/www/bejne/gaS9emiljad2Zs0 , aima.cs.berkeley.edu/ [3] www.csail.mit.edu/person/joshua-tenenbaum [4] soar.eecs.umich.edu/ [5] people.csail.mit.edu/phw/, ocw.mit.edu/courses/electrical-engineering-and-computer-science/6-034-artificial-intelligence-fall-2010/ - and many more... [6] dspace.mit.edu/handle/1721.1/67693, pdfs.semanticscholar.org/d95d/6e9d46572d04352563acbe84892fa9cd2b9d.pdf

[@xXJulensolo2Xx]

Couldn't you train a net to make small changes adversarial examples so they are recognizable but do no alter normal images, and then train another network with already processed image?

[@laurenswissels8480]

Awesome podcast, very thought ought questions!!!

[@shubhamchandra9258]

God of Deep Learning.

[@suhaylpatel9471]

Hi Lex, I recently discovered your channel and it has been really interesting. I am really considering learning more about this field in my free time, and coming from a mechanical engineering background, the only thing which is holding me back is the huge amounts of technical words that seem to confuse me more. Would you suggest diving straight into the theoretical side, or giving a go at some coding tasks that introduce me more into this field? I am undertaking a module in this but we've had no practical tasks to do, just theory and I've slightly neglected it due to the lack of application there was in the course.

[@VaShthestampede2]

1.) Sign up and complete Andrew Ng's "Deep Learning Specialization" on Coursera sponsored by DeepLearning.ai. 2.) Pick a small scope simple problem that you can at least conceptualize a way to use deep learning on, and write the simplest code you can in say TensorFlow or PyTorch to solve that problem. Or even, just reproduce an already solved problem using a framework. 3.) Be prepared to do lots of Googling and reading papers.

[@suhaylpatel9471]

@@VaShthestampede2 Thanks for the tips, I'll give that a go in the summer!

[@VaShthestampede2]

@@suhaylpatel9471 Andrew Ng is an amazing teacher. I'm also an ME, and his courses have helped me to at least be able to understand high level discussions.

[@chrisw6788]

Makes me wonder, how many years until ML produces a podcast like this on it's own and fools us all??? Great interview!

[@II_superluminal_II]

Chris W if I can buy a 50 gpu cluster, I could run a Conv3D GAN model (after sufficient time prepping the data of course) that can generate a “fake podcast”. No audio tho 🤣 that’s the only catch (audio+video is multimodal 😉)

[@antonyjames8133]

This is such a great conversation which makes you think (especially how Gans were made inspired from boltzmann machines). I wanted to ask does anyone know where this recording is taking place? As I can see people (probably computer scientists) are working and discussing in the background.

[@GodofStories]

lex says in the intro, it was done while Ian was at Google Brain

[@BlackHermit]

This guy is so cool! I wish I could be as cool as he is!

[@doruktopcu]

After this convo one can create anything with GAN, such as Ian's face for the rest of the interview using the first part.

[@sharang7858]

I was really really looking forward to this... :)

[@penguinista]

Humans don't learn pong by failing a million times, but humanity learned our common sense understanding of the world by our ancestors failing millions of times.

[@pandarzzz]

Thank you Ian & Lex for this informative video~! Wish you guys are well ✌

[@a1x45h]

Bruh I didn't notice he wasn't moving at the end

[@AbhishekKumar-mq1tt]

Thank u for this awesome video

[@MawaMaverick]

Thank you so much for this

[@kokopelli314]

Could learning optimization be applied to a brain model to improve human learning? Critical thinking, bayesian updating and differential and dimensional analysis all have that effect. Maybe ML can teach us to reason better.

[@blackrockcity]

Peter Diamandis proposed a prize to anyone that could improve learning efficiency in humans by 10x. It was a good idea but he never got it off of the ground. I think AI could put an Aristotle in our phone just as Steve Jobs envisioned decades ago. We tend to love the teachers that were most effective at communicating concepts (even if it turns out they were wrong because the science or history has evolved in the intervening years since your class). An AI could be so riveting as a teacher that it comes at the expense of other parts of your life like your job or relationship.

[@Silvertestrun]

Ty

[@princesingh9766]

please, Do the Podcast with George Hotz.(CEO of coma.ai )..

[@CE-vd2px]

He did

[@ahilanpalarajah3159]

Hi Lex, excellent interview thank you. Have you asked anybody about the future of differentiable programming? If not, could you please? Also, can neural ODEs enable us to provide prior information on a problem? Thanks in advance.

[@zeeshanashraf4502]

This is Gold!!

[@tiagoaoa]

He cites David Chalmers' Philosophical Zombies, nice..

[@andrewstang8590]

He hardly blinks his eye. I tried observing when he will blink his eye but had to blink mine. Ian is just super cool 😅.

[@irenef22]

Very clever disucssion. Impossible to listen to in the car...you should have max concentration on talk all the time. It is great, high value thing.

[@blakewisniewski3750]

I was here at 1k subscribers 😤

[@aspiceoflife]

1:05:02 says uh after saying he says no uhs

[@argrig]

For once the talk is on topic and lets the guest share his insight and expertise! It may get fewer views than the usual chit chat, but the intellectual value of such talk is infinitely larger.

[@user3sej90rga8]

Both of you guys are geniuses.

[@thevivekmathema]

he is a great guy. i am writing paper on gans

[@biswadeepsen3108]

Great Video!