Matt Botvinick: Neuroscience, Psychology, and AI at DeepMind | Lex Fridman Podcast
ฝัง
- เผยแพร่เมื่อ 12 มิ.ย. 2024
- Matt Botvinick is the Director of Neuroscience Research at DeepMind. He is a brilliant cross-disciplinary mind navigating effortlessly between cognitive psychology, computational neuroscience, and artificial intelligence.
Support this podcast by supporting our sponsors:
- The Jordan Harbinger Show: www.jordanharbinger.com/lex
- Magic Spoon: magicspoon.com/lex and use code LEX at checkout
EPISODE LINKS:
Matt's papers: scholar.google.com/citations?...
PODCAST INFO:
Podcast website:
lexfridman.com/podcast
Apple Podcasts:
apple.co/2lwqZIr
Spotify:
spoti.fi/2nEwCF8
RSS:
lexfridman.com/feed/podcast/
Full episodes playlist:
• Lex Fridman Podcast
Clips playlist:
• Lex Fridman Podcast Clips
OUTLINE:
0:00 - Introduction
3:29 - How much of the brain do we understand?
14:26 - Psychology
22:53 - The paradox of the human brain
32:23 - Cognition is a function of the environment
39:34 - Prefrontal cortex
53:27 - Information processing in the brain
1:00:11 - Meta-reinforcement learning
1:15:18 - Dopamine
1:19:01 - Neuroscience and AI research
1:23:37 - Human side of AI
1:39:56 - Dopamine and reinforcement learning
1:53:07 - Can we create an AI that a human can love?
CONNECT:
- Subscribe to this TH-cam channel
- Twitter: / lexfridman
- LinkedIn: / lexfridman
- Facebook: / lexfridmanpage
- Instagram: / lexfridman
- Medium: / lexfridman
- Support on Patreon: / lexfridman - วิทยาศาสตร์และเทคโนโลยี
I really enjoyed this conversation with Matt. Here's the outline:
0:00 - Introduction
3:29 - How much of the brain do we understand?
14:26 - Psychology
22:53 - The paradox of the human brain
32:23 - Cognition is a function of the environment
39:34 - Prefrontal cortex
53:27 - Information processing in the brain
1:00:11 - Meta-reinforcement learning
1:15:18 - Dopamine
1:19:01 - Neuroscience and AI research
1:23:37 - Human side of AI
1:39:56 - Dopamine and reinforcement learning
1:53:07 - Can we create an AI that a human can love?
Great podcast brother...
Try to bring Sam Harris if he wants to come. He doesn't want to do interviews these days. But you know how to ask meaningful questions. Best wishes for you man...
Post your old wrestling videos. Some of us are interested... i miss wrestling
Hi lex, I love your podcasts.
I had a professor of cognitive science at the university of California, Irvine by the name of Dr. Donald Hoffman. He is developing a theory of consciousness.
I think he’d be a great guest. He has a Ted talk and has been on the closer to truth series and many other podcast including Sam Harris not to long ago. Look him up I think you’d be very interested .
just for the sake of sound. get the reading of three vibrations of vocal patterns dharma meditation um sounds. under the domes in male, female and mix. then run billions of simulations through my old thumb drives. more than likely lost in a cafe or stolen by a friend.
@ 1:04:30 Fascinating, this is like saving the weights at an end of a trial/run, then clustering the state space for each new "box" (something similar was done in a 1997 thesis using Reinforcement Learning to control an inverted pendulum when going from simulation to an actual pendulum). I wish you can interview Sutton and Barto, Lex, they are really the "father's of RL", they do respond to emails (I emailed both of them and they did respond).
-John (from Arizona)
A conversation with Demis Hassabis would be very interesting.
🤩🤩🤩 You got that right
If possible, YES PLEASE!!!!
Couldn’t agree more!!!
It finally happened, it was interesting indeed.
Dear Lex, thank you so much for your wonderful podcast, during the past 2 years your videos have enlighten me to a level that i could not imagine. As a AI student, i have always learned from you, to be curious, and look at everything with an open mind.
It is great to see an AI researcher acknowledge that several of the fundamental works in Artificial Intelligence and neural networks was done in Psychology departments in U.S. Universities. Especially by Cognitive psychologists David Rumelhart, James McClelland and Joshua Tenenbaum. Also Walter Pitts, Warren McCulloch and Jerome Lettvin. This is wonderful. Thank you very much Lex Fridman and Matt Botvinick.
not tenenbaum, he was not even in high school at that time, i think Terrence srejnowski is a person you are missing, peter dayan, barto and sutton(sutton himself was trained in psychology) which allowed him to come up with the idea of temporal difference learning.
@@khwajawisal1220 There are many psychologists who contributed to AI - These include: Seymour Papert, Noam Chomsky, George Miller, Allan Collins and Ross Quillian, John Robert Anderson (ACT-R), Herbert Simon and Allen Newell, Yair Weiss and many many more. Tenenbaum did as well in the area of Machine Learning although that was in the late 90s and through the 2000s.
@@robinampipparampil Chomsky is not a psychologist. But yeah, his work was very important for psychology.
Please invite Josh Tenenbaum ! I'm pretty sure he just a few rooms across so you have no excuses
psychology is bs
Big fan of Dr. Botvinick’s work, especially his work on implementations of distributional rl in the brain and meta learning. Thanks for another great guest and conversation!
I believe an open mind is more important than a formally educated mind.
This is the first time I heard Matt Botvinick speak. He is a *great* communicator. Should have a podcast.
Thanks, I really enjoy the podcasts. Have you ever considered having two guests on at the same time?
You're killing it Lex. Love these conversations
Another very interesting conversation, thank you Lex!
Hi lex. AMAZING WORK . this chan is absolutely superb. Ive spread the word amongst all my software engineering friends. Truly great guests with intelligent questioning. No ads. Except start off. Very long shows. What more do u need. !TS A SUPERLATIVE SHOW!!!
this was one of the best interviews this guy has done! Interview technique is so much better, after every episode.
That ending was tear-jerkingly beautiful
Thank you Lex for the inspiring conversation! Your content has been fuel for my passion for AI
I love your interviews including your ads, even if I have to buy the things you recommend to assuage my deep feelings of intense guilt for misspelling your last name.
Thank you for putting on these interviews of so many extraordinarily great individuals.
Congratulations! One of the best videos of the podcast and on AI. Thank you so much!
Excellent episode, I really enjoyed this one!
Incredible interview! Thank you!!
This has given me so much motivation, given that I am currently a senior Psych major and that I am pursuing a Ph.D in neuroscience
its BS, also, read the unplugged alpha by richard cooper, human behaviour is derived from biology
I really noticed how, about 25 minutes in, Lex struck Matt's enthusiasm and humanity and he suddenly became really engaging to listen to. Nice job Lex.
Excited for this! Deep minds approach is simply awesome!!
This ties in nicely with two courses I'm currently taking. Thanks so much!
its BS, also, read the unplugged alpha by richard cooper, human behaviour is derived from biology
Wonderful again! MATT IS THE BEST!
That was awesome. What a humble and intellgent guy. I am a sucker for the virtuous cycle that is the engine of deepmind and other biologically inspired/constrained artifical intelligence work. Dr. Botvinick truly embodies this approach to understanding the mind/intelligence. Meta-learning via emergent rules, for me, was the most mind blowing part of the talk. A really, particularly, stunning example of complex behavior through simple rules.
Thank you so much for this one, Lex!
Big fan of yours Lex! Rogan and AI are my favorite podcasts out there. Stoked to see you are about to start skateboarding also. That pool in the background looks inviting. hahah. Keep it up.
Awesome podcast! Really sympathetic!
Not sure where this was filmed but I would be down to skateboard the pool in the background. Excellent conversation as usual, thank you.
After the Joscha Bach interview I can't help wondering what some of these other people have to say about his ideas.
It would also be interesting to hear comments about Penrose's models involving micro-tubules.
I'd also like less of the "what is the prefrontal cortex" and "do you see value in psychology" type questions that we can pretty easily get consensus opinions on without consulting the leaders in the field.
Sometimes the leaders' opinions are more valuable than any widespread consensus.
Intelligence is one of the structures of this Universe arranging it on so many levels. Things are simpler and make more sense. Love your work!
Absolutely love this. Thank you god for lex fridman
This is so wonderful!
"Flavor of champions" Sold !
Please interview Joshua Tenenbaum.
Nice podcast Lex!
Lex, you have a charmed life! Better than that is that
YOU created it. Bravo!
Amazing content!
With reference to the part near the end of the conversation about AGI needing a warm aspect to it in addition to an aspect of capability; I agree, we need to reward the friendliness, not only the efficacy. What is the warmth they refer to? I think it is; not only the absence of fear (of humiliation/ ridicule/shame/error/death) but also the building of self acceptance/celebration as a vital part within a system. We need to have AGI see itself & others as equal, yet gloriously different. We must reward respect. We must avoid punishing indifference, no? Thanks for this video.
Thank you Lex and Dr. Botnivick!!
While I was listening, I thought of a few common threads and would love to hear what people make of them.
On the topic of human/AI interaction and engaging with the possible consequences of AGI intelligence, it seems like a good analogy of our position as AI devs is that of an adult deciding whether to have a child. When making this decision, all parents realize that they will be responsible for the actions and development of their child as well as it's own happiness and health, yet we also accept that they are fundamentally beyond our total control. Taking this analogy even further, our approach to creating an AGI intelligence seems comparable to giving birth to an alien species - one that is not only beyond our control but that perceives/experiences the world, consciousness, and identity in a totally inconcievable way and on a different scale. We wouldn't even know how much of our own human perspective is transferrable to the AI, making us truly first time parents.
I know this metaphor of parent/child is pretty common in sci-fi dystopian literature/media, but i also think that understanding a non human identity relates to the last question of whether we will ever be able to love an AGI. Your response centered around the question of non-engineered and authentic "warmth" is really interesting and I would take it a step father to say that this prerequisite of warmth is actually tied to the identity of a system. For instance, while it is true that I love my parents, it is also true that I don't always respond to them with total warmth (i.e. investment in our talking and enthusiastic body language and tone). In fact, most of the daily interaction involved in a loving relationship can't be of this form, otherwise the "highs" would feel less of special. So beyond endless warmth, isn't what we're really looking for in AGI some perceived sense of genuine personality and developed identity? Without reaching a confident understanding of it's identity, it seems to me that we'll ever manage to develop feelings of intimate trust/love with an AGI, or be able to call it "a good guy".
What do y'all think?
Aliens arrive: “Take us to your..Lex”
14:40 ❤
"I fell in love with psychology and psychiatry with Freud when I was really Yung..." Ahaaahaaaha. Haha. Ha HaHa... Bonk.
Fantastic pun...
@@skierpage Both, but mostly Jung...he brought more into Consciousness.
Really Jung lol
its BS, also, read the unplugged alpha by richard cooper, human behaviour is derived from biology
22:42
"what's PDP?" that made me cry a little bit inside
Thanks for putting lots of efforts in making these podcasts. I am working to bring AI into space medicine. Your work motivates me to work harder.
Hey Lex, always a treat to experience these ‘talks’. With that said, especially regarding this show, it would be very interesting to have Dr. Donald D. Hoffman’s POV.
Also, it’d be cool to have a real jam session with you, in the spirit of love and possibilities. Thanks
Enjoyed!! Cheers! :-)!
This is outstanding. I was thinking about going back for a medical neuroscience degree to apply to computer science... I'm a senior software engineer now but I have this obsession with machine learning and that's what...sorry with artificial intelligence, and that's what drew me to you in the first place, Lex!
This is outstanding
8:25 great metaphor Lex. in response to matt's statement of psychology, cognitive sci, neuro sci unity. do an extra push up for me ;)
Thank you for an interesting interview; learning is on going, learning to work with other intellectuals across disciplines enhances our capacity to reach meta, knowing what humans want and delivering the kind of Society would be most ideal. Neural networking and AI may lead to a progressive future, " the greatest good for the greater majority", being in the middle actually gives a broader scope of perspective as one reconciles both worlds.
You have to get Peter Dayan, Yael Niv and Samuel Gershman!
i love the language of the both gentlmen. thank you for this revolutionary stuff. a truly bizarre era we are living in
Good questions Lex
This is the perfect podcast to listen to while working on a Leaky Integrate & Fire neuron simulation.
needs more blinker fluid.
Striking similarity of his voice with Lawrence Krauss
Hi Lex, really love your podcast. Please consider having Donald Hoffman on this podcast.
A conversation with Miguel Nicolelis would be very interesting.
Seems like we discount the information processing power of each neuron. Single cell organisms carry out surprisingly complex behaviors. A neuron takes in information from hundreds of sources and uses that information to determine whether and how to fire its axon. What factors go into that determination and how does that work at the molecular level?
People realize that real neurons are very complex and computationally powerful. A vast amount of research in neuroscience and in biophysics always went into studying specifically how single neurons work -- there has been several Nobel prizes for figuring out molecular-level mechanisms of their functioning. People in neuroscience use extremely detailed and accurate computational models of real neurons, which match well the behavior of the real ones. But this takes a lot of computing power -- such models include the actual geometry of the neuron with all its appendages, the distribution of different receptors and dozens if not hundreds of different kinds of ion channels, etc.
Top AI people (like Hinton or people at Deep Mind) are well aware of all this, and of the fact that artificial NN neurons are extremely simplified, comparing to the real ones. It would be more accurate to say that a single biological neuron is more similar to at least a small multiple-layer artificial NN, rather than to a single artificial neuron.
There are some AI researchers who use slightly more life-like models of neurons. For example, some make circuits with spiking neurons. But then they struggle to make them work as well as the more conventional models do. Unless one knew that there is some important feature in the real neurons that is really very helpful for making the circuits to work better, it seems unlikely that making artificial neurons more realistic would be advantageous. As far as I understand, that's the rationale for using such trivial and unrealistic models of neurons in artificial neural networks.
That's nice that engineers ask themselves the goal of their research. The problem I see, however, is that the company they work for can decide for them.
rate of neuron spiking needs much more time to estimate than the time scale of human reaction to events - it seems to be a combination of both rate and precise timing that carries the information depending on the circuit
Lex get John Carmack on!!! Let him talk some video games and vr!!!!
Lol I had your podcast on auto-play on here and thought you were talking to Lawrence Krauss the whole time
Cool vid.
I feel like you guys just need to talk to some sociologists and you'll make a bunch of breakthroughs, most of these hard questions about the world we should want to live in and how human groups interact have already been answered (to the extent they are answerable)
Can you recommend a couple of sociologists or books? I'd love to get into sociology a bit more
Yeah, good to know sociologists can tell us in which world "we" should want to live in.
Stop me if you've heard this before: I used to think that the human brain was the most fascinating part of the body, then I realised, look who's telling me that.
Thanks heaps for this, science graduate in psychology, inspiring me to work harder to get to that next level.
I just subscribed but i was amazed I hadn't before. Love your stuff, keep it up @Lex. One thing I thought of when Matt was talking about making an AI that can love or be "warm" is that... okay, yes, beating an opponent at a game can be seen as a goal, and it is only one dimensional. The thing is, we can imagine way more dimensions to playing a game, each with it's own reward distribution. So my question is, could an empathetic dimension be realized in this game context as for instance: letting the opponent win sometimes. Also, that seems to be part of a "play" behavior in mammals and other animals which includes multiple sub-dimensional rewards like, "the want to keep playing". I guess, empathy and thusly other emotions end up posing an unanswered question along the lines of: Is it possible for humans to perform truly selfless actions? I mean, I know that people do things for others, etc. but we always do get that dopamine reward even if we cannot measure a material reward for helping others (sometimes to our own deficit). What does it mean to "be warm" from a logic perspective?
@Lex, I have a question for you. Before that, thank you so much for doing this. Matt layed out the top down and bottom up approaches in AI-Cogntion-Neuroscience. What do you personally believe? To give you my opinion, as Joscha said in your other interview, if there are things being built then let them build it. Even François for that matter said, L5 is something which is very hard to achieve and they are sticking to the vision, as in the deep learning part of it to begin with and I believe, as he mentioned in his paper, it is going to be system centric generalization at first (hopefully) anyway. What if the bottom up approach can contribute effectively to this mutually exclusive domains in broader terms. What do you personally believe or think considering you staying at the AI end of it?
A conversation with peter thiel would be interesting
I'd love to hear you talk with Ed Boyden.
1:12:30 What he is describing might function as a definition of intelligence. Measuring the capacity of meta-learning capabilities of a system may act as an indication/metric of intelligence.
I'd love to see Denny Britz
I heard about another botvinik, he had huge influence on post ww2 soviet era of chess
The analogy between dopamine and reinforcement learning is pretty jaw dropping
Idk where I would be today if not for neuroscience podcast
As someone from Computer Science background and wants to contribute in the field of AI, is pursuing higher education in computational neuroscience helpful?
Fix your site Magic Spoon! I want it!
Do you think there is an infinite (or extremely large) descent kind of analysis required in neuroscience? In physics we can always reason all the way down to the atoms, but in neuroscience we are trying to explain why certain combinations of atoms are acting in concert. This transforms a problem space from just the physical space into a larger "combinatorial" space (2^(Number of Atoms in a Brain))
Please do part 2 and go crazy on the romanticized and philosophy side 🙌🏼
Naftali Tishby would be great to see @Lex Fridman.
I Saw an ad seconds after you finished your intro ad, after the talk had started, and just saw another around 14-15 minutes in... I really hate ads while watching something, and totally don't mind yours at the beginning, but not while the conversation is going on... (don't know if there's more yet... but seems likely..)
Fascinating discussion. How about making the first AGI a politician? (not to set the bar too low tho ;)
The best cereal is plain old porridge. Ask any Scot. 😊👍
Ok, Lex FrYdman!
So much hot air!
Maybe my second favorite conversion, after the one with Lex's dad.
Is that fence around the pool to keep skateboarders out?
I recommend the book Gesture and Speech by Leroi-Gourhan, page 154 " cerebral evolution of the neantropics " , Is an anthropological and also archaeological point of view, he introduce the prefrontal cortex as the main reason for hour conscience. I also recommend the work of Searle and his research on the Conscience , he is a pioneer to bringing the biological study of conscience in the academy system
Lex, why don't you have a Discord server?
1:28:15 about a minute before he said this I started playing a Beethoven sonata.
Please do podcast with Josiah Zayner
Injective protocol a very smart project brings us excitement to always follow. I believe @InjectiveLabs project will work and succeed because this project is with a great team that always provide creative ideas and those ideas make this project unique.
@InjectiveLabs #InjectiveProtocol #investing
You should get Bernardo Kastrup on your podcast.
john carmack!
Baby raptor crossing at 45:10
RE: building "warm" AI. Getting answers to how "warm" an AI is will be exceedingly slow compared to getting answers about how proficient it is at a computational task. We can choose the latter in an algorithm but in the former a human being has to be in the loop. Getting better at making warm AI system necessarily means making systems that are about to generalize from less data. I think AI research is well on its way!
Have Ken Stanley on the podcast
How does dopamine affect sleep dreams?
Lex, if you do one with Tim Urban or Max Hodak i'll build you a statue.
Will the statute be capable of having the podcast that will hold me over until the Kurzweil and/or maybe the Downstreamers from Stephen Baxter’s work - great podcast so I heard but the buffering on the video is taking way too long.
In my program they already teach dopamine as reward prediction error as an accepted truth 😂
Which program is that?
@@zulya007 currently just finishing up my first year of a PhD in the ELSC Theoretical and Computational neuroscience program at The Hebrew University of Jerusalem.
I actually have an exam on Tuesday and dopamine in the Basal Gangalia is a topic on the exam.
1:17:39 Hello mystery person in blue
1:43:43 AH! They're inside now lol
Invite Richard Stallman and Linus Torvalds for a point of view from FLOSS. Also Jaron Lanier may be interesting.
Last time I was this early, psychology and neuroscience were still different things.
55:00
I already love my iPhone, in a way, so why not an AI that loves me to flourish ;)