I really enjoyed this conversation with Tommy Poggio... from Einstein and time travel to the mystery of the visual cortex, compositionality of nature, neuroscience of ethics, and Flowers for Algernon.
Thank you Lex! This interview is amazing and very valuable. Your questions were perfect and let Tomaso share with us all his incredible amount of knowledge, and wisdom
It is hard to imagine how could you see through the complexity of life and technology to bring out so many great questions that many people have in mind but fail to formulate, and it is greatly delightful to hear that these fundemental questions answered or discussed that triggers many deep thoughts in science, in life and in ourselves. Thanks for sharing, these knowledge and insights are the good fortune of our generation!
My attempt at outlining this conversation at the granularity of questions. Note that sometimes professor's answer steered in a different direction than expected by the question, so entries in the outline are based on answers. 0:0 - introduction 0:52 - what fascinates Poggio the most about Einstein's genius ? 2:08 - what was special about Einstein's thinking ? 3:39 - possibility of time travel 4:45 - the dream which inspired Poggio to study intelligence 6:16 - AI as a tool for solving great mysteries of nature 7:43 - problem of human intelligence 8:55 - can we build AI without understanding the brain ? 13:07 - comparison between artificial and biological neural networks 14:50 - what is missing the most in artificial neural networks ? 16:53 - hardwiring vs learning in brain 20:12 - educated guess about hardwiring vs learning 22:36 - modularity of the brain 25:32 - are functions of modules hardwired or assigned during early learning ? 27:58 - basic questions we don't know about the brain 29:13 - levels of abstraction needed to study brain 30:45 - how much lower levels help understanding higher levels ? 32:53 - compositionality in neural networks 35:20 - are problems solvable by deep neural networks necessarily compositional ? 39:18 - learning algorithms 42:43 - why do simple algorithms work quite well ? 44:48 - universal approximation theorem 45:48 - what about the size of approximating network ? 47:27 - how deepness helps approximation 47:54 - how important are GANs in the roadmap to general intelligence 51:15 - joke about history of CS 52:16 - importance of motion in vision 55:05 - difference between recognizing objects and understanding the scene 56:20 - existential threat of AI 58:05 - when will AGI be reached ? 59:01 - will AGI system be explainable/understandable ? 1:00:40 - understanding in a sense of building another copy of AGI 1:02:40 - ethics of neuroscience vs neuroscience of ethics 1:04:14 - is ethics hardwired ? 1:05:45 - problem of consciousness 1:07:10 - is consciousness required for AGI in a sense of self-awareness 1:08:30 - is fear of death required for consciousness/intelligence ? 1:09:50 - next breakthrough 1:10:32 - visual intelligence. Experiment about sense of self 1:13:08 - what does it take to be successful in science/engineering ? 1:15:16 - advice for researchers 1:16:40 - role of arguments 1:17:47 - one question for a genie 1:18:23 - flowers for Algernon. Is intelligence a gift ?
Back in 2008 summer I was a freshman working in Prof. Poggio's lab at MIT. I was interested in AI but this was before the time of deep learning, the lab was working on object recognition system inspired by the visual pathway. I was tasked with building an image collection/labeling system (before the time of image net) to train the system...my supervisor told me after this first project I would start working on real AI stuff. I remember when Prof. Poggio coming into the office one day, he was kind and asked how I was doing. I had no idea how big of a deal he was (everyone was at MIT), and just made polite small talk. This is the conversation I wish I had back then with him. Then I would've stayed at the lab instead of leaving just before the cusp of deep learning, disillusioned by the tedious data collection process.
I would vote for a spoiler alert about "Flowers for Algernon" at 1:18:30 Such a great book and much more effective if the reader doesn't know how the plot proceeds. BTW, I love your work, Lex. You are one of my favorite interviewers. It has been astonishing to watch your skills progress. Excellent series.
33:07 Are you meditating on the question @Lex Fridman? :-) It is very fascinating how you take care that you get the question right. Tomaso's silence on the last question is more intriguing than any argument or answer to that question.
This was a fascinating conversation on many levels. I found Thomaso to be the ideal mentor I would like to have, openess, curiosity, ambition and fun in the journey of understanding and discovery. It would be great to have his points summarized, he shed some great light on a lot of questions. I specially enjoyed the ones on how we are beginning to understand NN, like the high probabilties of finding a global minima in a highly paramatrized model due to high number of prameters and the fact that hidden layers somehow solve the need for having N ^ Dimensions parameters in order to approximate the function, wasn't completely sure I understood if N was the error percentage boundary that you want or if 10 is a rule and 10% error was just a boundary. Anyway thanks to you both.
So wonderful to have an expert so openly admit they didn't know something... (right at the end) even while being bigged-up, and pause for deep thought. Respect 👍
Amazing conversation with my hero Professor Poggio, great questions followed by thoughtful answers. With collective approach and minds across vast earth we will accelerate our understanding but as Prof. Poggio said, lets have an open mind and embrace the different ideas to have a chance solve the mystery. Fun and Curiosity. Thank you Lex for another great interview.
You can tell Tomaso is an incredibly wise person because he admits he doesn't know, and most of the discussion was Tomaso going away from all kinds of speculations, he is a fact guy, he either knows or doesn't know. And for most of the metaphysical question he doesn't know what the answer is probably nobody knows.
Simple, so yeah intelligence it's making conscious of the logic that governs us and management better. Love its just attraction, by pure logic you love anything, you have it or not you will fear losing it or not having it and you will hate or defend what you love... Love, fear, and hate are perfectly connected... Buda teaches this that's why he says, detachment, it's the way to end suffering. You can program a robot or computer to emulate this emotion and behavior. Love its attraction, Fear its defend, Hate its attack.
You could see the burden of his answer in his demeanor; it said a lot about the man. Suggested to me at least that he's seen the dark side of 'intelligence' and 'genius' - here's a man without illusions!
7:20 - why I transferred from biotech to comp sci, AI (if we can make it human level or greater) will potentially be the fastest route to the greatest rates of technological development.
- electrical engineers creating computer components could be likened to neurologists and neurosciencentists on a low level - IT professionals could be equated to a psycologist or psychiatrist, helping us as the users to troubleshoot the software - what we are looking to do here as programmers is to bridge that gap, creating an emulator for the whole system, needing to understand both sides of the mind and how they fit together There is not yet an equivalent to emulator programmer for the human mind to help us programmers understand this connection.
Interesting question in the end. Intelligence, and happiness. If intelligence also consists question of emotional attachment. as a human is a social animal, happiness and intelligence also mean the ability to connect with people on different levels. but not even smart person can do so (not every smart person know how to build a deep relationship). and if think in this direction, if AGI has to also have this ability to make a connection with people. and how it would work.
Love its just attraction, by pure logic you love anything, you have it or not you will fear losing it or not having it and you will hate or defend what you love... Love, fear, and hate are perfectly connected... Buda teaches this that's why he says, detachment, it's the way to end suffering. You can easily program a robot or computer to emulate this emotion and behavior. Love its attraction, Fear its defend, Hate its attack. Simple, so yeah intelligence it's making conscious of this and management better. You can suffer by loving fake invented human stuff like social status or love status or if your car its clean or not.
Subjectivity could be the nucleus to AGI. What's missing in technology today is subjectivity, not the definition of intelligence. If the robot is subjected to GI then it will become AGI. Sofia the robot is a perfect example, I'm sure she will exhibit AGI eventually.
If a human adult level conscious machine can be created, it will have to be based on the human brain. Gerald Edelman has done the heavy lifting and provided the blueprint in his book, The Remembered Present, and other works. The people at the Neurorobotics Lab at UC at Irvine, and their future students, will probably be the ones to realize that blueprint.
I found your question r/g "compositionality" at around 35 minutes into the video very interesting ....where Tegmark takes it as 'atoms/quarks/the four forces' are "real" and Poggio takes it as 'we evolved to survive via compositionlizing the physical universe' but sadly neither Poggio or Tegmark ever question whether we're still basing our conclusions on wrong assumptions and thus why it's all so confusing and not 'fitting' to what we observe - the "flat earth or geocentric syndrome". Wonder why these two (or many other very intelligent and gifted scientists) don't question - could it be that reality is some form of a computation from within which our brains evolved to compositionalize/simulate a world of people/places/things/space....thus all these things/parameters (including atoms/quarks/etc) are not what the brains see/observe but actually CREATE as some form of simulated version of things/people/places/space. That to me is the only explanation of how the brains do what they do. There's no way in the world where neurons firing or not firing (regardless of how 'complex' the networks be) can RE-create the 'real external physical world' within the skull other than the more likely scenario of the brain's computation "CREATING" the simulated external virtual world (similar to how it does that during our dreams). I think people like Joscha Back and Steven Wolfram are closer to what's going on then unfortunately some of these other geniuses who are still stuck on trying to explain all this while believing very strongly on the existence of a 'real' physical world out there .....not much different than very intelligent 'flat earthers' trying to come up with deep explanations on why we can't find the edge of the earth or why it doesn't go on forever!
How does that change the discussion though? Whether the world is simulated or there exists a ding ang sich out there we still have to contend with objects [artifacts?] in the sub-[ob]jective space that will kill you and end both your simulation and its supposed* reality. When my motorcycle collides with an SUV the fundamental organization of the colliding forces have the same value as the taste of the sun! Is your point one of credulity?
Better pricing of error; a child learns 'this is fire!' once - every other repetition is penalized heavily; machines have gone their route because they cannot be punished [suffer] - when we engineer pain receptors the machine learning curve will go vertical
Problem of intelliegence. origin of life. Origin of the universe Solve all the problems. Key to intelligence Gradient descent - adjust parameters Reinforcement learning. Polynomial s
The problem with Max Tegmark's argument is also that he believes his idea about the world as physics describes it, is the reality itself. Bur it may really narrow down to the fact we describe reality on those terms, in mathematical language based around the principle of equilibrium, etc, because this is what our brain allows us to do. It might be that if we had a "different kind" of brain, we would discover a different kind of "mathematics", that is to say a symbolic logic based on different principles, which are unimaginable to us. The common belief amongst many physicists that scientific equations are the description of reality can not be supported logically (because of extrapolation of Godel paradox). This is a kind of hubris a neuroscientist, or even Lacanian psychologist can easily point out. There is a difference between what's Real(which we never directly encounter), and our Symbolic imagination about it. When you approach the foundation of physics with clear mind you must accept that ideas such as Force or Energy are metaphysical ideas, which function only within a linguistic structures of our minds, and are not found in nature. Every physicist should know the difference between a physical phenomenon, and the description of that phenomenon. Nobody have ever observed a force or an energy, only, as we are used to put it - the effects of force and energy. But to claim World is an effect of something (like energy of singularity) is itself a philosophical a'priori idea, coming directly from Newtonian a'priori definition f a Force. It's a very simple bit of reasoning which all physicists almost deliberately refuse to accept. The map is not the territory, no matter how good your map is. And let's be frank, physics account for a minuscule part of world phenomena, they simply decided to call everything they know nothing about "noise", or "chaos". It's a really cheap trick.
@@skierpage You could have been kinder but my thoughts exactly. Of course better omelets may be made with dinosaur eggs - get back to us when you serve one!
Can neural nets interact with each other like different areas of the brain do and as well as different glands, and gut microbiome interact with the brain? i mean 90％ of the body's serotnin is found in the gut, 8％ in the blood, and a whopping 2％ is found in the brain. Something tells me psychiatry offices could benefit from an in building blood lab. Did you know about the autoregulators in the brain? So by testing the blood, psychiatrists should be able to get a better understanding of how much serotnin a person needs. the only thing is they are only beginning to understand how SSRIs affect the brain. there are 100s of neuro transmitters all interactinteracting with different parts of the brain as either a neurotransmitter or a hormone depending on where in the brain and body they are and what situation are in and what you are trying to accomplish at the time. Please watch Stanford University's Lecture Series by Robert Sapolsy. he has the grooviest lectures on human behaviorology. And one on Advanced Neuroscience.
This is why time travel is so hard - the first civilization who win all wars in the universe realized time machine is ultimate weapon to end all wars as Avenger Endgame suggested, they decided to go to the roots of everyone else to make sure no one else can ever make a time machine. Just kidding! Another conspiracy no one can ever prove and disprove.
Alcoholic lip. Dad had that, too. i suspect this AI generated. i don't remember ever subscribing to lex fridman's channel. Imagine my surprise when i noticed subscribed in grey.
