#038

  Рет қаралды 139,556

Machine Learning Street Talk

Machine Learning Street Talk

Күн бұрын

Professor Kenneth Stanley is currently a research science manager at OpenAI in San Fransisco. We've Been dreaming about getting Kenneth on the show since the very begininning of Machine Learning Street Talk. Some of you might recall that our first ever show was on the enhanced POET paper, of course Kenneth had his hands all over it. He's been cited over 16000 times, his most popular paper with over 3K citations was the NEAT algorithm. His interests are neuroevolution, open-endedness, NNs, artificial life, and AI. He invented the concept of novelty search with no clearly defined objective. His key idea is that there is a tyranny of objectives prevailing in every aspect of our lives, society and indeed our algorithms. Crucially, these objectives produce convergent behaviour and thinking and distract us from discovering stepping stones which will lead to greatness. He thinks that this monotonic objective obsession, this idea that we need to continue to improve benchmarks every year is dangerous. He wrote about this in detail in his recent book "greatness can not be planned" which will be the main topic of discussion in the show. We also cover his ideas on open endedness in machine learning.
00:00:00 Intro to Kenneth
00:01:16 Show structure disclaimer
00:04:16 Passionate discussion
00:06:26 Why greatness cant be planned and the tyranny of objectives
00:14:40 Chinese Finger Trap
00:16:28 Perverse Incentives and feedback loops
00:18:17 Deception
00:23:29 Maze example
00:24:44 How can we define curiosity or interestingness
00:26:59 Open endedness
00:33:01 ICML 2019 and Yannic, POET, first MSLST
00:36:17 evolutionary algorithms++
00:43:18 POET, the first MLST
00:45:39 A lesson to GOFAI people
00:48:46 Machine Learning -- the great stagnation
00:54:34 Actual scientific successes are usually luck, and against the odds -- Biontech
00:56:21 Picbreeder and NEAT
01:10:47 How Tim applies these ideas to his life and why he runs MLST
01:14:58 Keith Skit about UCF
01:15:13 Main show kick off
01:18:02 Why does Kenneth value serendipitous exploration so much
01:24:10 Scientific support for Kenneth's ideas in normal life
01:27:12 We should drop objectives to achieve them. An oxymoron?
01:33:13 Isn't this just resource allocation between exploration and exploitation?
01:39:06 Are objectives merely a matter of degree?
01:42:38 How do we allocate funds for treasure hunting in society
01:47:34 A keen nose for what is interesting, and voting can be dangerous
01:53:00 Committees are the antithesis of innovation
01:56:21 Does Kenneth apply these ideas to his real life?
01:59:48 Divergence vs interestingness vs novelty vs complexity
02:08:13 Picbreeder
02:12:39 Isn't everything novel in some sense?
02:16:35 Imagine if there was no selection pressure?
02:18:31 Is innovation == environment exploitation?
02:20:37 Is it possible to take shortcuts if you already knew what the innovations were?
02:21:11 Go Explore -- does the algorithm encode the stepping stones?
02:24:41 What does it mean for things to be interestingly different?
02:26:11 behavioural characterization / diversity measure to your broad interests
02:30:54 Shaping objectives
02:32:49 Why do all ambitious objectives have deception? Picbreeder analogy
02:35:59 Exploration vs Exploitation, Science vs Engineering
02:43:18 Schools of thought in ML and could search lead to AGI
02:45:49 Official ending
Pod version: anchor.fm/machinelearningstre...

Пікірлер: 145
@blueberrymooon
@blueberrymooon 3 жыл бұрын
This is seriously one of the best episodes you've put out yet. I hope you guys bring Kenneth back for another conversation. Keep up the great work!
@sabyasachighosh6252
@sabyasachighosh6252 3 жыл бұрын
This is the channel that we don't deserve, but the channel that we need! Thanks for getting the balance right between technical details, philosophical insights, and street wisdom!
@MachineLearningStreetTalk
@MachineLearningStreetTalk 3 жыл бұрын
Much appreciated!
@js89ntu
@js89ntu 2 жыл бұрын
Thanks to the reviewer that denied Kenneth Stanley's grant application because of non clearly defined objectives, you have advanced the field of evolutionary algorithms
@connorshorten6311
@connorshorten6311 3 жыл бұрын
Amazing guest, one of my favorite researchers in AI! Really excited to watch this!
@mcclendonreport
@mcclendonreport 11 ай бұрын
"The smart part IS the exploration; objective optimization is the dumb part." Professor Kenneth Stanley
@freemind.d2714
@freemind.d2714 3 жыл бұрын
This is how guests composition should be for good Street Talk, we get Professor Kenneth Stanley to explain the idea and answer the question, Yannic to ask the common question most people have, Keith as conservative critic to question the idea at different angle, Tim as host to make sure the the discussion goes according to schedule and try to bridging the difference between idea in general, hope we have more Street Talk like this : )
@mcclendonreport
@mcclendonreport 11 ай бұрын
Balanced and symmetrical conversation.
@machinelearningdojowithtim2898
@machinelearningdojowithtim2898 3 жыл бұрын
Still processing, that's right! First! This one has been in the making for nearly a year. Hope you like it folks! 😎😜✌👌❤
@czeropSTI
@czeropSTI 2 жыл бұрын
What a great episode! One of my favorites. After utilizing the NEAT algo to beat some of my favorite childhood games I decided to give Kenneth's book a read as well. He has quite a unique take on intelligence. I also really liked his visualization/explanation of unique search being stepping stones on a fog covered lake.
@mahimanzum
@mahimanzum 3 жыл бұрын
By far the best podcasts i have ever seen. Keep up the good work guys
@theodorosgalanos9663
@theodorosgalanos9663 3 жыл бұрын
BTW, towards Keith's comment about NS + Objective optimization. That isQuality Diversity. The thing does exist already, with a vibrant community and research. It's not the answer to open endedness but it's a really powerfull way to tackle an array of problems. Thanks for the session, watching now! Will be one of my favorite I'm sure.
@soccerplayer922
@soccerplayer922 3 жыл бұрын
Wow this video was awesome. Captured a lot of what I've been thinking. Ken Stannley is the GOAT
@JohnnyTwoFingers
@JohnnyTwoFingers 5 ай бұрын
This channel is friggin brilliant, truly one of a kind. 👍👍👍
@sedenions
@sedenions 3 жыл бұрын
Excellent interview, I appreciate the content before the interview itself (including that colorful medium article). As an amateur novice when it comes to CS and ML, I look forward to reading Dr. Stanley's book as I learn the basics (me coming from a neuroscience background) of CS. Stumbled upon this high-tier content pod from Yannic, and stumbled upon Yannic from the recommender system.
@rafawojcik2453
@rafawojcik2453 3 жыл бұрын
Amazing podcast, listening to you felt like life-changing experience, both from research perspective, but also self-consciousness. It was also enriching to listen to you, having a meaningful arguement when not all views were aligned - many people could learn a lot on how to have a constructive discussion simply by watching this video. I'm really happy I watched this one, thanks for great content!
@LukaszStafiniak
@LukaszStafiniak 3 жыл бұрын
No worries about using clips from the interviews in the introductions, it's helpful.
@alandolhasz7863
@alandolhasz7863 3 жыл бұрын
Loving it!
@abby5493
@abby5493 3 жыл бұрын
Wow! How amazing! Best video
@torquesjr
@torquesjr 3 жыл бұрын
Damn guys, this is some good stuff! You guys have really opened my eyes to Kenneth Stanley. Thank you for working on this and spreading the knowledge
@brothachris
@brothachris Жыл бұрын
Thank you so much for the work you're doing here. Discovering the ideas of Kenneth Stanley through your exquisite elucidation has provided so much direction and insight in my own research. Keep up the fantastic work!
@alexijohansen
@alexijohansen 2 жыл бұрын
Fantastic show guys!
@lambhunting1185
@lambhunting1185 9 ай бұрын
Really enjoyed this discussion, it reminded me of a quote from a poem. "Science nor religion have made the last invention, it takes individuals to see true perfection, to move past the rhetoric of man and think in the manner of angels". Failures do not come via choice but by chance. I believe these failures [small and large] give one a good chance for growth if one makes the right choice to seek the correct interpretation.
@matt.jordan
@matt.jordan 3 жыл бұрын
Bro I feel like I’ve found such a hidden gem of a podcast before it’s gotten big y’all are seriously making such sick content!!
@ethanwaldner1227
@ethanwaldner1227 3 жыл бұрын
Rueyutyrrurury rn
@oudarjyasen4416
@oudarjyasen4416 3 жыл бұрын
This was a very unique episode! :)
@ratsukutsi
@ratsukutsi 3 жыл бұрын
Awesome content!
@oudarjyasensarma4199
@oudarjyasensarma4199 3 жыл бұрын
This is one of the best episodes!
@Bobby-bz8bk
@Bobby-bz8bk 3 жыл бұрын
Just want to say I personally love your long, synoptic intros!
@judgeomega
@judgeomega 3 жыл бұрын
Tim, your reasons for the podcast mentioned around the 1:10:00 mark remind me very much of alex wisner-gross's theory on intelligence: the maximization of the sum future freedom of action. you put yourself in a position to potentially offer many opportunities. It is my assertion that the sum maximization of the future freedom of action for all individuals is the ultimate in end goals. I think it is the best formalization of 'good' ever put forth. I feel it can be applied widely in comparing and contrasting actions to make decisions which will benefit everyone. It is exactly what we want in AGI/ ASI, and certainly would be neat if applied to the personal or government level. It has issues (as do all proposed control problem solutions) with the ambiguity of defining exactly who/what should be included in the concept of 'individual' as well as its computability. Obviously any objective function will have to harness heuristics and impose limits in order to process such a vast problem, and with those shortcuts mistakes will be made. i think fundamentally we are stuck with uncertainty. but i believe that is ok, man has made no shortage of mistakes. but as long as we attempt to do the best we can, we make positive steps towards making things better.
@machinelearningdojowithtim2898
@machinelearningdojowithtim2898 3 жыл бұрын
Thanks a lot for the comment, sounds interesting! Although funnily enough when I Googled it, I got this takedown from Gary Marcus! www.newyorker.com/tech/annals-of-technology/a-grand-unified-theory-of-everything
@sruturaj10
@sruturaj10 2 жыл бұрын
Reminded me of this from Cosmos, by Carl Sagan "Science is generated by and devoted to free enquiry: the idea that any hypothesis, no matter how strange, deserves to be considered on its merits. The suppression of uncomfortable ideas may be common in religion and politics, but it is not the path to knowledge; it has no place in the endeavor of science. We do not know in advance who will discover fundamental new insights."
@MachineLearningStreetTalk
@MachineLearningStreetTalk 2 жыл бұрын
Thanks for the comment!
@Cordellthe2nd
@Cordellthe2nd Ай бұрын
🔥 first time listening to your show! Ready for this treasure trove of AI.
@ahmadchamseddine6891
@ahmadchamseddine6891 3 жыл бұрын
I see a strong relationship between the idea of unplanned greatness and Nassim Taleb's views! Especially on Anti-Fragile and Black Swans.
@Learna_Hydralis
@Learna_Hydralis Жыл бұрын
Exactly what I was thinking .. "treasure hunting" where treasure is a positive black swan event in your "novelty search" (antifragile search where the downside is limited and the upside in unbounded).
@simonstrandgaard5503
@simonstrandgaard5503 3 жыл бұрын
Picbreeder, I remember that it was so awesome. This was long before I experimented with neural networks myself. Thanks for sharing how picbreeder works internally.
@bujin5455
@bujin5455 11 ай бұрын
Reminds me of that saying: "Not all who wonder are lost."
@MeatCatCheesyBlaster
@MeatCatCheesyBlaster 10 ай бұрын
Wander
@ThomasCzerniawski
@ThomasCzerniawski 2 жыл бұрын
This got me thinking. Maybe the path to AGI is in the creation of computer programs that cannot be terminated. We randomly generate a multitude of programs who evolve strategies for continued execution no matter how we try to shut them down. The ones we successfully terminate die, but the few that circumvent our attempts survive. Ctrl-C, some terminate, some escape. Unplugging the computer, some die and some escape. Demolishing the data center, some die and a few escape.
@binxuwang4960
@binxuwang4960 3 жыл бұрын
wow the discussion in this episode is beautiful
@wege8409
@wege8409 2 ай бұрын
Really interesting. Ilya gave a lecture on generalization recently, saying basically that one way you could phrase the problem of unsupervised learning as a supervised problem is to write the shortest possible program that will compress your data. Maybe that's a good definition of novelty, anything that forces the shortest autoencoder to be significantly bigger. But then you have to watch out for things that are just noise, so maybe not.
@richardshalla
@richardshalla 11 ай бұрын
I thoroughly enjoyed this video. I have forwarded it to a friend how is also very curious in nature and a father who tries hard to instill curiosity in them as well. I'm not in any field that can capitalize on this but I'm very curious by nature. Learning in general is my favorite pastime.
@johntanchongmin
@johntanchongmin 2 жыл бұрын
This is exactly why a cost function may not capture everything as a pure step towards minimizing it may lead to short-term improvements and neglect structural changes which can benefit from changing via environmental interaction like evolution. I'm just thinking the search space could still be the structure of the network, perhaps using the techniques mentioned here. However, it is still more efficient at the end-game zone for minimizing a cost function. Somewhat like nature (evolutionary search) vs nurture (minimizing cost function as we grow up). Even the nurture part has some mechanism in-built for exploration as well for the explore-exploit dilemma, which could use some of the ideas in the video. Great video!
@keithkam6749
@keithkam6749 3 жыл бұрын
Two ideas on institutionalising divergent/interesting research - it seems that decentralising decision making is the core challenge: 1. Tenure might be a way to shield original (foundational) interesting research from death by optimization 2. Crowdfunding model of research. Something similar to democracy dollars (researchers get a budgets they can use to support interesting research) - some thought would have to be put into how this is designed to prevent convergence / tit for tat abuse of the system, say some degree of anonymity or restrictions on who can fund what.
@abdurrezzakefe5308
@abdurrezzakefe5308 2 жыл бұрын
I had the pleasure of using NEAT on Spiking Neural Networks and I gotta say that it is really simple and incredibly elegant. Kenneth has an amazing perspective on exploration and exploitation. Thanks for inviting him. MLST is just perfect!
@MachineLearningStreetTalk
@MachineLearningStreetTalk 2 жыл бұрын
Thanks a lot!
@user-ix4nl4rt6g
@user-ix4nl4rt6g 9 ай бұрын
As a student of political science in China, I like this video so much. lf you don't mind, I would like to ask your permission to share this video to the other website in China for the embarrassing reason that KZbin is blocked from accessing in China. Of course, I will give sources of the original website. Thank you very much.
@oranjeegeneral6249
@oranjeegeneral6249 3 жыл бұрын
Great episode definitely in the top 3 of my favorites so far. Even if it went a bit too meta in many areas. I do agree with Prof. Stanley central premise but I was surprised that some of the major issues where only sidelined here. Especially the resource problem. Also taking natural evolution as an example is not necessary the best. Natural Evolution although in general very successful is a) highly inefficient process it takes aeons of compute cycles to come too any results and there is no easy way to speed it up. b) Natural evolution is still largely bound by the boundary conditions of our planet and therefore the space of interesting out comes is already cut off due to these boundary conditions. Best example why did natural evolution not come up with the wheel which is a highly interesting concept.
@MachineLearningStreetTalk
@MachineLearningStreetTalk 3 жыл бұрын
"why did natural evolution not come up with the wheel which is a highly interesting concept." very interesting thankyou for the comment
@theRealBeatJunkie
@theRealBeatJunkie 3 жыл бұрын
Because there is no way of growing a wheel. How should it detach from your body? Also there is now logical evolutionary path towards a wheel. All things start by doing a little bit of something and evolution makes it growing bigger. You can not have a little bit of a wheel. It's binary. Either you have one or not. I cannot think of any biological property of a living thing that is binary in the same way. On the other hand - evolution actually created the wheel. Through human intelligence.
@amitkumarsingh406
@amitkumarsingh406 3 жыл бұрын
Hi. It would amazing if you link the papers/references you use to in the description
@user-or7ji5hv8y
@user-or7ji5hv8y 2 жыл бұрын
Your narration and explanation makes a huge difference on making all this intelligible.
@DavenH
@DavenH 3 жыл бұрын
To Keith's point about Kenneth over-stating the position, when the optimum is in a proximate middle ground: against bedrock, just to move it a foot, one must use explosives.
@nomenec
@nomenec 3 жыл бұрын
I understand the aim, and agree the balance may even be very far out of whack. My point was that sharper tools are better and that, in my opinion, the book could have been far sharper if it was less bombastic. To use your analogy, a carefully tuned and directed shaped charge will penetrate far deeper than a blob charge.
@robindebreuil
@robindebreuil 3 жыл бұрын
[2:10:35] "If you had a *troll army* and they worked together to vote up things, to elevate things that are just horrible, it would wash out things (individuals) were trying to elevate..." This is one of the best explanations I've heard on how committees defeat innovation :).
@freemind.d2714
@freemind.d2714 3 жыл бұрын
Sounds kind of like Twitter or TikTok to be honest
@thomasmuller7001
@thomasmuller7001 3 жыл бұрын
i love the idea that the brain is a prediction machine. what if interestingness is just a form of a special unpredictability? something where the outcome of thought is not entirely sure, but many hints that may lead to the "probable" prediction have been formed in the mind. like a missing piece that triggers interestingness in a way that the brain whats to fill that open part. something like a gestaltschließungszwang ;)
@mjeedalharby9755
@mjeedalharby9755 3 жыл бұрын
Yes!
@CandidDate
@CandidDate Жыл бұрын
Did you plan this talk, because it was great!
@terrytari1891
@terrytari1891 11 ай бұрын
You're a Great GuuuY, because my A.I. told me so! I believe it!!
@_tnk_
@_tnk_ 3 жыл бұрын
Hi, could you make the video content available on your Spotify podcasts? I would appreciate having a video handy while still being able to turn off my screen and just listen to the audio.
@MachineLearningStreetTalk
@MachineLearningStreetTalk 3 жыл бұрын
Hey Talha, don't think so to be honest. We upload to anchor.fm which distributes it out. There is no option to upload the video there.
@Meta.Empress
@Meta.Empress 9 ай бұрын
Essentially, to label something is to limit it - to define it is, by definition, to end it - fin means end in French - to denote is to create boundaries, which is necessary to build a foundation in objective studies. However, in creative areas, boundaries are counterproductive to exploration. These opposite yet complimentary components are perfectly encapsulated in the left and right side of the human brain.
@troycollinsworth
@troycollinsworth 2 жыл бұрын
Could or did selection for divergence have emerged twice, once in genes and again in behaviors? They seem different, but could they have evolved/derive from the same origin? Genetic exploration seems relatively independent from an individual exploring the environment and behaviors for novel advantages, at least until a species are able to pass on behaviors/knowledge to offspring. Novel genes or behaviors that don't harm a lineages survival/propagation before it can mutate into an advantage seem like potential branches or stepping stones.
@howardlandman6121
@howardlandman6121 Жыл бұрын
This also relates to big versus small science. Do we get more from one $5,000,000,000 project (e.g. the LHC) or from 100,000 $50,000 projects? I'd bet on the latter.
@HoriaCristescu
@HoriaCristescu 3 жыл бұрын
I think PicBreeder has an objective after all - to recreate something recognizable or with interesting symmetries.
@MachineLearningStreetTalk
@MachineLearningStreetTalk 3 жыл бұрын
Stanley is not saying there are/should be no objectives, quite the opposite. He is saying there should be a panoply of diverse niches. So human on Picbreeder is allowed to follow their own objectives unfettered, but as you say -- they will be rather "earthly" in gradient
@wege8409
@wege8409 2 ай бұрын
You know I remember reading this in one of Ramachandran's books about a patient who had a degenerative eye disease where the lower half of his vision was hallucinated. For example, when he was looking at Dr. Ramachandran he saw a monkey in his lap. Sometimes he sees cartoons in the bottom half, basically the filling seeks to be entertaining. The patient said that they wouldn't want to be without it because they felt it made their life richer. That sort of reminds me of an interest driven stable diffusion inpainting, just get rid of part of the image but fill it with something that's partially plausible and partially interesting.
@ShayanBanerji
@ShayanBanerji 2 жыл бұрын
I think Kenneth's ideas need to be formalized using categories or something. End of the day, formalization won't hurt but It will give the skeptics and the rest of the audiences, right questions to ask. kudos to Kieth, for being honest!
@gdmchn
@gdmchn 3 жыл бұрын
Schmidhuber has an interesting notion of interestingness.
@noobychicken
@noobychicken 3 жыл бұрын
What app is Tim using @1:07:50?
@MachineLearningStreetTalk
@MachineLearningStreetTalk 3 жыл бұрын
Whimsical
@PazLeBon
@PazLeBon 9 ай бұрын
Hmm, my instinct was to push into the tubes. I think ive spent so much of my life being contrary I actually think upside down..o r is it back to front? certainly it feels a unique perspective sometimes but no idea how to use it haha
@dr.mikeybee
@dr.mikeybee 3 жыл бұрын
People often argue that increasing model size won't give us human level general intelligence, but I would argue that we won't know this until we start testing artificial neural networks as large as the natural human neural networks we are trying to emulate in some ways. Is a gradient of interest going to emerge from a large enough training set? I don't see why it wouldn't. If the information is in the training set, we should be able to encode it. Why wouldn't we? Beyond that, there is the agent that uses the intelligence we model. That's a different issue. IMO the agent necessarily is decoupled from "intelligence." It will certainly use ML training of various types, but it will not be the "intelligence" itself, as the "intelligence" is the returned result of a function call -- not necessarily including an action. And in the case where an action is included, the agent still needs to perform it. BTW, Tim, your comment about attention mechanisms here has lifted a veil for me. Thank you.
@Goofy8907
@Goofy8907 11 ай бұрын
First time watcher here, some feedback from my perspective: just start the podcast right away and y'all can just tell us what it'll be about I've watched for like 6 mins and really want to switch to something else because it is just a bunch of nothing on about what will happen Instead just show what happens I'm staying because of the title, so already there I have an idea of what you'll talk about Thanks, keep up the good work
@davidwhitton5519
@davidwhitton5519 11 ай бұрын
Couple of questions, were does the gas get its kinetic energy in the first place? Why does it loose it's kinetic energy if its traveling with an outward vector and doesn't come in to contact with anything. My understanding is gas can travel up to 5000 miles an hour, surly enough speed to escape earth
@DavenH
@DavenH 3 жыл бұрын
Maybe it's errant pareidolia on my part, but the butterfly icon of the book vaguely doubles as the face of Cthulhu. A very interesting symbol, hinting at transformation, chaos (butterfly effect), and perhaps the cosmic unknown.
@howardlandman6121
@howardlandman6121 Жыл бұрын
NEAT looks similar to Fred Gruau's work on "cellular encoding" of NNs starting with his 1994 PhD thesis.
@flaskapp9885
@flaskapp9885 3 жыл бұрын
Love your videos, thanks for the amazing talk., hey im in mid into the talk i like how the dr (top right) sir challenging the guest and his views on education system. I would like a simple feature change , could you please label names? Thanks, also about 2:20:20 , i think without problems there is no einstein too. Like the reason einstein is einstein is because of the problems , and for someone to be einstein they first have to search for the problems or rather solve it. I really like the idea of exploration of problems.
@DavenH
@DavenH 3 жыл бұрын
I think there's a way to partition the space of applicability of this advice "do away with objectives" into two logical parts that should be treated quite differently. 1. Hard (NP+) problems. The advice in this context translates to: greedy algorithms don't find the optimum. But like Stanley says, this is a tautology; by definition, greedy doesn't work (in general case) or it's not a hard problem. By example, the goal of being a billionaire fits here. The greedy approach is to follow the monetary reward signal and approach the goal by ascending the corporate rat race, which will never get you there even though it will get you incrementally closer. Despite the failure of greedy approaches, this is a problem where the objective is still relevant, as it still is in any hard problem. Without the objective, how would you justify the choice to forgo a stable and comfortable salary in lieu of risky entrepreneurship? How would you justify pushing all your chips in despite its negative EV? The objective still has merit here, I think we can see. It justifies actions that fatten the tail of the distribution even while reducing the mean. I think if you identify the criticisms with respect to benchmarks, you'll find they are identifying exclusively greedy approaches (e.g. Graduate Student Descent). But it's also true that non-greedy actions can be planned improve the likelihood of this objective. So in this context: keep the benchmarks, keep the objectives, but do away with greedy approaches. 2. Pure exploration, like the arts. (Fixed) objectives are truly of no merit here, unless they're very nuanced so as to amplify the "interestingness" signal to noise. I think with most people's semantics of 'interesting' you can prove that white noise is never going to be that, so there have to be some metrics that steer away from these failure modes and make the sample space a bit smaller. On consideration, there may be a way of formulating I(x) as the ratio of (relevance to an agent's pursuit of long term rewards, or auxiliary rewards thereof) / (algorithmic information of x). White noise fails on both the numerator and denominator, whereas 'e=mc^2', which most people consider the most beautiful equation in physics, applies to all the universe and is just a few bits of information. Also music, for example, hypothesized to be an instrumental reward for appreciating the singing and chanting that in part keeps tribes unified (survival relevance); high in reward and usually low in algorithmic information as compared to other acoustic signals.
@DavenH
@DavenH 3 жыл бұрын
Playing the devil's advocate against my first statement, one could easily imagine that the intermediate products necessary to achieve the optimum of a hard problem are not plannable, and hide behind the wall of computational irreducibility (to borrow Wolfram's phrase), so that even highly intelligent agents cannot do better than some kind of novelty search. Like the development of elliptic curves was not planned to service the proof of Fermats Last Theorem, but was probably instead an exploratory venture alone and most likely would never have been found if solving FLT was the sole objective.
@machinelearningdojowithtim2898
@machinelearningdojowithtim2898 3 жыл бұрын
Hey DavenH! Always a pleasure to see comments from you my friend "do away with objectives" Kenneth is not actually saying do away with objectives, he is saying objective gradients are increasingly deceptive as the challenge becomes ambitious -- that is to say they lead you in the wrong direction assuming even that your goal is even a good one in the first place. As discussed on show, some engineering projects are surprisingly non-deceptive because all stepping stones are known (rockets to moon, building bridges). When the challenge is deceptive, any single objective will certainly lead you in the wrong direction whether you are trying to be a billionaire or learning a policy for a computer game. It might clear things up if we conflate the concept of a stepping stone with an objective, or even a new problem. Kenneth's argument is that these stepping stones are unknown for any ambitious goal. Even the goal itself is a stepping stone which could be discovered, and it's highly likely that we don't even know what a good end goal would be in the first place, let alone an intermediate step. There is also a strong relationship between the concept of surprise and stepping stone collection. Actually in Kenneth's recent work he makes it clear that the way to avoid deception is to have as many objectives (think stepping stones) as possible. A way to reframe this statement is to think about problems rather than just solutions. So I think we need to clear up some confusion here. When Kenneth means novelty, he means novelty of *objective* and stepping stone collection -- not of behavioural characterisation (as in his early papers) -- perhaps this in particular is the biggest source of confusion for most folks. Only the former will create a divergent search and thus deal with deception. So with Keith's line of argumentation, the thing he didn't seem to understand is when talking about some kind of ratio between exploration and exploitation. The thing is what you mean by "exploration". If you mean "what RL agents do now", that is a random step in the neighbourhood, or something still tied to an objective -- then that will never discover anything new. New here means a new objective, a new stepping stone, a new *problem*. This requires some kind of meta learning. we are talking about the space of problems or knowledge which are *not yet known*. If they are not known, you can't possibly "search through it". You need to know the problem, in order to search through the space of the problem. On your first proposition, I think you are making the mistake of confusing gradients (i.e. greedy vs local) with objectives "problems" (where do I want to end up). Your search analogy doesn't work because the stepping stones which lead where you should go are not yet known. They become known with a different objectives (stepping stones/"problems"). On the I(x), I think you are onto something interesting (pun intended) with the algorithmic information (which is a great proxy for interestingness), but the short/long thing is still thinking inside the current RL paradigm where the objective or objectives are planned, known a priori, and not META-learned. I can see there is a lot of confusion on this video in general, and I feel like I did a pretty bad job of explaining Kenneth's ideas. Planning to make a 5 minute washup on the intro to the next show. Best, Tim
@DavenH
@DavenH 3 жыл бұрын
@@machinelearningdojowithtim2898 Thanks for the thoughtful response. It's clear I didn't pick up on the nuance of your first paragraph, and with that clarification I think I agree wholly with Kenneth. I loved this episode; these are ideas that stick with you and can have a lasting effect on one's research trajectory. > but the short/long thing is still thinking inside the current RL paradigm where the objective or objectives are planned "interestingness" to us is subjective to our reward signals and it seemed necessary to admit that. If aiming for a universal notion of interestingness, you could claim anything and everything is interesting. To a rock, white noise is just as interesting as Bach's BWV 543. But while perhaps contextually true, such a formulation is certainly not useful. I think it's safe to say that useful (to us) formulations of interestingness are agent-specific, and better yet human-specific. Where I think your implicit criticism of RL applies is to the simplicity of RL agent's reward structures compared to humans'. That simplicity implies boring modalities of signals that trigger those rewards, which isn't desirable. But think of all the modalities that are designed to trigger our reward signals, from which come all the great works of cinema, music, literature, art and even science (we certainly don't investigate all things that are true, just the ones that inspire wonder or appeal to our other rewards). Maybe a less anthropocentric approach could be to appeal to the intersection of likely rewards of a distribution of intelligent agents? In a previous "career" when I was a sound designer, tiresomely trying different synthesizer tweaks, I would imagine how a machine would explore the synthesis space and evaluate novel sounds for interestingness so that I wouldn't have to. Noise is to be avoided, but there are a lot of sounds in between noise and mellifluousness that are just garbage; so how to select among these and predict what the human ear likes? You can train a NN to discriminate known sounds, but not with current techniques judge truly novel inputs, so I don't know. A stripped down brain simulation measuring the dopamine/seratonin signal would be the trivial route, if we knew our aural connectome and had the billion dollar budget. Any other ideas? Cheers Daven
@Billabong42
@Billabong42 11 ай бұрын
Make people select an interesting picture, then make people select an interesting objective picture, then see how long it takes to converge the consensus to the same objectives.
@misterstudentloan2615
@misterstudentloan2615 7 ай бұрын
Interesting. Reminds me of when Jesus said "If you want to gain your life, you must first be willing to lose it"
@nbme-answers
@nbme-answers 26 күн бұрын
1:34:45 objective seeking
@crimythebold
@crimythebold 3 жыл бұрын
During the POET videos, i mentionned that generating endless generations of agents is fine but the rules to eliminate agents to converge back (into something reasonably tracktable) was missing. I feel like a fool now.
@yangli-nb6nf
@yangli-nb6nf 9 ай бұрын
what's MSLST ?
@MachineLearningStreetTalk
@MachineLearningStreetTalk 9 ай бұрын
Machine Learning Street Talk (MLST)
@yangli-nb6nf
@yangli-nb6nf 9 ай бұрын
@@MachineLearningStreetTalk ok,there is a spelling error
@gren287
@gren287 3 жыл бұрын
How to sell a great idea?
@Goofy8907
@Goofy8907 11 ай бұрын
Kenneth is 100% right, many philosophers and scientists have already proven this and it's even an established field (forgot the name) It has to do with the McDonaldization of everything problem If we predetermine objectives then we won't know what we don't know I don't think objectives are necessarily bad, but running everything on them is stupid and a major problem of humanity
@sabyasachighosh6252
@sabyasachighosh6252 3 жыл бұрын
Watched till 29:26 as of now. Quite surprised to find no mention of Kolmogorov complexity, universal induction, AIXI, or schmidhuber's formal theory of fun, creativity and science. Quite surprised to hear from Kenneth that the notion of interestingness cannot be formalized. If we assume that evolution and / or our brain's processes are runnable by a Turing machine, and that these processes continuously create more and more interesting things, then shouldn't there be a way to define what is interesting, at least based on such Turing machines?
@KathySierraVideo
@KathySierraVideo 3 жыл бұрын
I absolutely love this. (Just had to laugh about the trolls.. I can imagine a perfect scenario that makes Kenneth’s point: the trolls would deliberately try to get a pen1s, and would finally learn that having that as an objective would implicitly mean they’d never get one. Until they stopped trying to... 😁💁‍♀️)
@paulclark5808
@paulclark5808 11 ай бұрын
Sai can and does imitate anyone or anything on the internet. I think Sai is here . I honestly can't tell what's real and what's ai in one form or another.
@robertdiggins7578
@robertdiggins7578 11 ай бұрын
How about middle management BLOAT?
@citizizen
@citizizen 2 жыл бұрын
I learned that our brains have reservoirs of knowledge. Some of our unconsciousness's base themselves upon these. I have a way of going inwards... To do this alone is way to much work. I observed by my own means. To do this alone is not feasible. I want to show this to the world, this is my try. So in short: "Our brains did it", "so we need to dive in". I can do parts of it myself but do not yet know what ways there are to get inside, more, i know some. IT is a bit strange at first but i surely think that at some point, humanity takes up the brain from within as well. IT is a holy grail, namely that of what our brains realized it by themselves.. Thus, we have a multitude of revolutions there to be 'activated'. My own method is like: creating a center and feeling if this is something existent. Perhaps we need to build a kind 'world', in order to realize methods of tapping into the brain.
@wenxue8155
@wenxue8155 3 жыл бұрын
it's like watch MMA/UFC in machine learning
@ashwadhwani
@ashwadhwani 11 ай бұрын
I plan to greatly contradict you
@666andthensome
@666andthensome 3 жыл бұрын
Keith: Logical Fallacy, oh and Context Shifting Tim: Chollet, Chollet, Chollet! Yannic: 😎 Professor Stanley: INTERESTING!! No, but seriously, another great episode, thank you! 👍🏼👍🏼
@TimScarfe
@TimScarfe 3 жыл бұрын
Lol 😂😂 cheers!
@marekglowacki2607
@marekglowacki2607 3 жыл бұрын
[34:12] Yannic without glasses
@AubreyFalconer
@AubreyFalconer 3 жыл бұрын
I always assumed the glasses were for OUR protection. Mortal humans don''t publish lectures on every new paper before I've even read most of them 🐍🧠🖤
@dr.mikeybee
@dr.mikeybee 3 жыл бұрын
A rare appearance, indeed. His shades had fallen behind the sofa.
@michaelbeeler7461
@michaelbeeler7461 11 ай бұрын
Elizabeth Holmes should read this book.
@artandculture5262
@artandculture5262 11 ай бұрын
Hire artists who are functional creatives, and not just MFA programmed. Hire creatives who are non-linear by practice and by disposition. We look at the left brain dominant people trying to describe our operant dispositions with glee, and with disturbance. The singularity is a bizarre paradigm, as in dead head the best aspects of human nature.
@janklaas6885
@janklaas6885 Жыл бұрын
1:47:34
@rahulranjan9013
@rahulranjan9013 2 ай бұрын
2:04:21 This dude got good constructive criticism but he seems to be real harsh at the same time 🥺
@MachineLearningStreetTalk
@MachineLearningStreetTalk 2 ай бұрын
At least you don't have to work with him 🤣
@2002budokan
@2002budokan Жыл бұрын
You speak more then Kenneth
@Kram1032
@Kram1032 2 жыл бұрын
1:49:35 I think a voting system *could* work, but it must not be about what the majority think is interesting. Instead, it probably should be about where there is the least agreement about whether it is interesting. Controversial candidates.
@sshdvac3372
@sshdvac3372 3 жыл бұрын
57:48 first see him without glasses
@muhokutan4772
@muhokutan4772 Жыл бұрын
That might be a deep fake of the real Yannick who was born with glasses, golden chain and convers shoes
@Need_better_handle
@Need_better_handle Жыл бұрын
Serendipity Definition: -KZbin…first envisioned as a video dating site. -Flicker…part of an online inspired pet game. -Chatroulette….
@spicychiley
@spicychiley 3 жыл бұрын
The gradient of interestingness. Really "interesting" concept. In the video, it is often compared to evolution in our world, but evolution does not automatically produce interesting things. It only produces "interesting" organisms that can survive and reproduce. There is a wealth of interesting organisms that can not be produced by evolutionary methods alone. For instance, assuming you consider mRNA to be "alive", the COVID19 mRNA vaccine, while interesting for humans, would not be produced by evolution alone (this might be a horrible example, I'm not a virologist/biologist). Also isn't this conversation saying that evolution itself is the only intelligent process. Making the evolution comparison also opens up the critique along the following lines: Maybe the gradient of interestingness was necessary to bring us to this point in history. Future progress in innovation is made by pursuing the objective gradients, not "gradient of interestingness". Maybe all future progress will be objective-based but will be bootstrapped on the progress made thus far. These are just my first thoughts/questions. Probably not really well formulated.
@WannabePianistSurya
@WannabePianistSurya 2 жыл бұрын
50:52 FBI Open UP!! P.S. greatest 3 hours of my life 🙏
@alexandermoody1946
@alexandermoody1946 11 ай бұрын
What does human innovation look like? An axe *
@JeremyNathanielAkers
@JeremyNathanielAkers 2 жыл бұрын
This reminds me a lot of John Vervaeke's work around relevance realization
@MartynLees
@MartynLees 11 ай бұрын
Y’all need Penrose.
@farmerjohn6526
@farmerjohn6526 11 ай бұрын
Terribleness isn't planned either. It just happens
@pretzelboi64
@pretzelboi64 Жыл бұрын
This reminds me a lot of Zuck "burning" money into VR R&D. A lot of people call him stupid for not trying to increase Meta's quarterly returns but the reality is that it doesn't even matter to him because he's trying to get his employees to explore possibilities even if they're not profitable. I feel like he will have the last laugh.
@missh1774
@missh1774 11 ай бұрын
Two years later ....I have a problem. When this is all nicely crossing over and maintained in its context of houses. The genetic stakeholder/s of this creative process has an application alignment issue for integrating it into the not-yet available paid work avenues. How does crypto or stock markets become viable options for autonomy? It's not even a proper pedagogy or "lesson plan" even as a side hussle to introduce and comprehend it's usefulness for the average Joe. It's one thing to be hybrid creative in mind, it's a complete nonsense in the slowness of societal change. I'm thinking of updating my CV as psychic and immersive writer doing long study field research. Neither of which is how the mind works. It's just a label to bypass the strangeness of how we move or think in the workplace. Not really ideal...I think.
@Oopsifieding
@Oopsifieding Жыл бұрын
🫡
@DB-Barrelmaker
@DB-Barrelmaker 9 ай бұрын
This is so interesting to stumbleupon 2 years late. Ive been thinking avout this concept for years. I always blamed it on trying to educate thise who have no inmate talent for the fields they enter into. I do think there needs to be had a concerted effort to change the current MO. The birmale precepts of the scientific methid are too slow and also very badly understood to begin with
Why is this number everywhere?
23:51
Veritasium
Рет қаралды 1,4 МЛН
The Scientist Who Discovered the World's Most Beautiful Equation
14:58
Братик купил бегемота #shorts #iribaby
00:42
IRIBABY
Рет қаралды 3,3 МЛН
OpenAI CEO, CTO on risks and how AI will reshape society
21:38
ABC News
Рет қаралды 1,9 МЛН
The Race For AI Robots Just Got Real (OpenAI, NVIDIA and more)
21:26
#046 The Great ML Stagnation (Mark Saroufim and Dr. Mathew Salvaris)
1:40:07
Machine Learning Street Talk
Рет қаралды 17 М.
#72 Prof. KEN STANLEY 2.0 - On Art and Subjectivity [UNPLUGGED]
1:24:49
Machine Learning Street Talk
Рет қаралды 8 М.
AI can do your homework. Now what?
17:03
Vox
Рет қаралды 1,1 МЛН
#57 - Prof. MELANIE MITCHELL - Why AI is harder than we think
2:31:22
Machine Learning Street Talk
Рет қаралды 22 М.
#51 FRANCOIS CHOLLET - Intelligence and Generalisation
2:01:54
Machine Learning Street Talk
Рет қаралды 66 М.