Karl Friston's Unfalsifiable Free Energy Principle

Подписаться 46 тыс.

Просмотров 11 тыс.

50% 1

World-renowned neuroscientist Karl Friston discusses his unfalsifiable Free Energy Principle, a formal account of variational free energy minimization via Bayesian methods between systems and environments. According to this principle, the brain is a predictive engine that is constantly trying to make accurate predictions about the world in order to minimize its own uncertainty. We cover technical details, surprisal, Markov blankets, and much more.
❶ Free Energy Principle Overview
3:22 What is the Free Energy Principle?
11:30 What's the origin story of the Free Energy Principle?
15:45 What's the scope of this discovery? Scale invariance, individuation, basal cognition...
21:50 Definition of "free energy"
25:15 Information = Energy
26:15 F.E.P. dual to Jaynes' maximum entropy principle
27:50 Preferences are built into Bayesian beliefs within F.E.P.
32:35 Is the brain minimizing free energy because it's resourceful/greedy?
35:13 Variational free energy = accuracy - complexity
39:10 Expected free energy = expected accuracy - expected complexity
40:40 Bayes optimal: expected information gain (maximize) & expected cost (minimize).... perception is just hypothesis testing
45:45 The free energy principle conforms to itself!
❷ Surprisal Minimization
47:10 Does a system want 0 surprise?
50:45 The dark room paradox
54:40 Intuiting maximizing information gain and minimizing surprise
1:00:15 E.T. Jaynes (maximize entropy | measurements & beliefs) vs Erwin Schrodinger (minimize entropy | homeostasis & outcomes)
1:01:50 What about winning the lottery?
1:04:20 The Twilight Zone | A Nice Place to Visit episode
1:08:50 2 ways to minimize prediction error: 1) Change the model 2) Act on the world
1:12:25 How do we decide which of the 2 to do? Larger mammals take turns...
1:14:53 Seems like every organism samples about 4x per second (4 hertz)... attenuation/saccadic suppression
1:19:45 So that's why you can't tickle yourself!
1:22:00 If you weren't able to properly attenuate (e.g. Parkinson's disease)
1:24:00 Anecdote from Carlos (fire escape, police chase)
❸ Markov blankets all the way down
1:27:40 Allan Hobson would remind you about dreaming
1:28:20 Non-lucid dreaming & Markov blankets
1:31:45 Is the environment reciprocating or inferring the system's internal states?
1:33:00 The environment is learning about you
1:33:50 "Realize your environment is composed of other things very much like you."
1:40:55 Dissolving the exploration-exploitation dilemma (EETO)
1:45:00 The Marriage Problem (game theory)
1:47:20 Hamilton's Principle of Least Action
1:51:35 Jeremy England's dissipative adaptation (thermodynamics)
1:54:15 "You won't get more fundamental than the free energy principle"
1:56:30 The complexification that seems inevitable (Susanne Still)
1:58:13 Is the universe a quantum error correcting code?
2:01:05 Thoughts on Donald Hoffman's work? Holography
2:03:40 Is there an outside? Philosophical issue
2:05:37 What is active inference?
2:08:20 Free energy principle simulates active inference | The only thing you get from the F.E.P...
2:09:40 If you could give your 20-year-old self one piece of advice?
🚩Links to Karl Friston 🚩
The free-energy principle: a unified brain theory? (Friston 2010)
www.uab.edu/medicine/cinl/ima...
The free energy principle made simpler but not too simple (Friston 2022)
arxiv.org/pdf/2201.06387.pdf
Homepage: www.fil.ion.ucl.ac.uk/~karl/
Karl Friston: Active inference and artificial curiosity presentation (2017)
• Karl Friston: Active i...
🚾 Works Cited
The Twilight Zone "A Nice Place to Visit" 1960
Truncated 5-minute version: • Twilight Zone Lite: A ...
Midjourney, miscellaneous images
Every Life is on Fire (Jeremy England) www.amazon.com/Every-Life-Fir...
🚀 What is this channel?
Exploring Truth in philosophy, science, & art. We'll uncover concepts from psychology, mythology, spirituality, literature, media, and more. If you like Lex Fridman or Curt Jaimungal, you'll love this educational channel.
p.s. Please subscribe! Young channel here. =)
#science #complexity #surprise #friston #attractors #brain #surprisal #neuro #neuroscience #cybernetics #activeinference #bayes #bayesian

Опубликовано:

25 апр 2023

Ссылка:

Скачать:

Готовим ссылку...

Добавить в:

Мой плейлист

Посмотреть позже

Комментарии : 59

@Carlos.Explains Год назад

❶ Free Energy Principle Overview 3:22 What is the Free Energy Principle? 11:30 What's the origin story of the Free Energy Principle? 15:45 What's the scope of this discovery? Scale invariance, individuation, basal cognition... 21:50 Definition of "free energy" 25:15 Information = Energy 26:15 F.E.P. dual to Jaynes' maximum entropy principle 27:50 Preferences are built into Bayesian beliefs within F.E.P. 32:35 Is the brain minimizing free energy because it's resourceful/greedy? 35:13 Variational free energy = accuracy - complexity 39:10 Expected free energy = expected accuracy - expected complexity 40:40 Bayes optimal: expected information gain (maximize) & expected cost (minimize).... perception is just hypothesis testing 45:45 The free energy principle conforms to itself! ❷ Surprisal Minimization 47:10 Does a system want 0 surprise? 50:45 The dark room paradox 54:40 Intuiting maximizing information gain and minimizing surprise 1:00:15 E.T. Jaynes (maximize entropy | measurements & beliefs) vs Erwin Schrodinger (minimize entropy | homeostasis & outcomes) 1:01:50 What about winning the lottery? 1:04:20 The Twilight Zone | A Nice Place to Visit episode 1:08:50 2 ways to minimize prediction error: 1) Change the model 2) Act on the world 1:12:25 How do we decide which of the 2 to do? Larger mammals take turns... 1:14:53 Seems like every organism samples about 4x per second (4 hertz)... attenuation/saccadic suppression 1:19:45 So that's why you can't tickle yourself! 1:22:00 If you weren't able to properly attenuate (e.g. Parkinson's disease) 1:24:00 Anecdote from Carlos (fire escape, police chase) ❸ Markov blankets all the way down 1:27:40 Allan Hobson would remind you about dreaming 1:28:20 Non-lucid dreaming & Markov blankets 1:31:45 Is the environment reciprocating or inferring the system's internal states? 1:33:00 The environment is learning about you 1:33:50 "Realize your environment is composed of other things very much like you." 1:40:55 Dissolving the exploration-exploitation dilemma (EETO) 1:45:00 The Marriage Problem (game theory) 1:47:20 Hamilton's Principle of Least Action 1:51:35 Jeremy England's dissipative adaptation (thermodynamics) 1:54:15 "You won't get more fundamental than the free energy principle" 1:56:30 The complexification that seems inevitable (Susanne Still) 1:58:13 Is the universe a quantum error correcting code? 2:01:05 Thoughts on Donald Hoffman's work? Holography 2:03:40 Is there an outside? Philosophical issue 2:05:37 What is active inference? 2:08:20 Free energy principle simulates active inference | The only thing you get from the F.E.P... 2:09:40 If you could give your 20-year-old self one piece of advice?

@TheIgnoramus Год назад

I’ve been in multiple rabbit holes recently and this is just icing on the cake.

@Jacob011 Год назад

01:57:20 This is probably the most important point for those wondering about the peripheries of FEP. I'm glad Karl confirms what I thought all along. You assume you have a blanket to begin with.

@brynoreilly2731 10 месяцев назад

What a wonderful discussion - thank you! Of all the videos that I've watched with Professor Friston, I'd say he was at his most natural here. Palpating my own senses, there even seemed to be a hint of his eccentric vulnerability that came through, which was nice to 'see'. You no doubt had the effect of bringing that out in him.

@williamjmccartan8879 Год назад

About 38 minutes in, it sounds like the lowest cost of free energy is achieved through passive transformation. Wonderful way to finish the conversation gentlemen.

@elitequeef Год назад

Never a disappointing lecture Carlos! Been reading Sheldrake's stuff on Morphic Resonance. Cheers and stay blessed sir.

@Carlos.Explains Год назад

Thank you AB! I will check out morphic resonance... first time hearing about it. :)

@MikeCampo Год назад

@@Carlos.Explains would really love a talk with Sheldrake on his morphic resonance ideas! I first heard him discuss it in 2014 on the Joe Rogan podcast of all places and it's very fascinating. Would be interesting to hear how its progressed. I wouldn't be surprised if there's a connection to the work Michael Levin is doing, which I'm finally digging into and loving.

@kittyhinkle3739 11 месяцев назад

I think Morphic resonance describes why the orcas are tipping boats quite well.

@xmathmanx 7 месяцев назад

morphic resonance was a good theory but now we have michael levin's work showing that bioelectricity, more so than DNA, is the main driver of morphology

@MadplowDisease 11 месяцев назад

These podcasts are amazing great job!

@atrocitasinterfector Год назад

Dude this guy is awesome

@elarah08 Год назад

@1:25, Carlos telling the story of the car being chased by a police car unexpectedly, where he observed his brain constructing reality - I can confirm that many times: I am very myopic, nearly 6 diopter, and one day decided to no longer use my glasses. First I saw nearly nothing, only blurry areas, I always was afraid to get lost... but then my brain adapted, it actually learned to make hypotheses of blurry areas in some distance - they changed, when I walked towards the specific objects in question. For example I walked to a crossroad, something was standing there. First I saw a dog with a human. The breed of the dog changed 2 times. Then the picture became blurry and it developed to a person with a trolly - pictures constantly changed over and another reality (with details!!) emerged into form. And I was permanently moving my head, to get different angles for my eyes, which I also only observed (did not plan for those movements). When I approached the scene I saw it was actually a woman with a bag standing beside her. It was a fascinating experience to watch my brain developing a new form of sight, adapting to the new blurry data-input. Today I can see surprisingly sharp - but that's an illusion. Today I can see some trees perfectly in a distance, which I missed very much for 2 years (their beauty) - even the leaves are moving in wind with sparkling light - but then suddenly there appears a big blurry obstacle between me and the tree - which then turns out to be a person or some object. I get very startled and nearly jump and scream, this is how unexpected the object appears in my vision. So I know, I don't SEE the farer away trees so sharply, my mind knows how a close-up from a tree looks like and projects it on to the scenery, where it has identified blurry objects as trees. Same with house-fassades. It fills in details to make my vision richer, which also calms me down - because I have become rather self-conscious in regards of my sight and orientation. But till today I have the similar effect you described, that now and then I see how my reality gets changed in front of my eyes, so to say. Amazing experience, and it changed my whole way of thinking, of "looking at things". We are permanently told what we CANNOT do, and those are the limitations we accept in our reality - literally. But it's simply not true. I am very alone with my observations, nobody is as insane as I have been to take off my badly needed optical aid - till you told the story of the 2 cars in this video - which is why I am so exited and write such a long comment 🥴 Fascinating talk with Prof. Friston, thx for that.

@zebleckDAMM 9 месяцев назад

that sounds amazing, i want to experience that wow

@scottjones-singersongwrite6193 8 месяцев назад

Wonderful stuff!

@NightmareCourtPictures Год назад

Energy fundamentally would be what we experience of a system undergoing change. So when things stop changing, then there is no more "energy" to take from it...But this is only true if you take that the universe isn't infinite and not scale invariant...which it probably is (that the universe is scale free and will continue creating more complexity forever). If that is the case then there will always be energy, at all scales as systems come to maximal entropy, and there will always be enthalpic processes with it. Those two properties together (scale invariance and infinite universe) means there will always be free energy...rather that energy "is free" because there will always be change...and therefor there is always a way for observers at any scale, at any complexity to undergo enthalpic and entropic processes. You can paint a more visual picture of this by just going into game of life...creating some random initial conditions, and then coarse grain the game by blurring it a couple times (by a factor of 4x - 10x) You will notice that the automatons, look like blobs at the coarse grained level...and these black blobs will behave almost like solid objects...but notably jiggling around a bit. But the most important observation is that, it's not that the black blobs are what constitutes as having energy, that is a human construct...it is the DIFFERENCE between the white gaps, and the black blobs. the perceived energy (jiggling) only exists to an observer when both of those things are present...and therefor it is the difference, and how those differences change through time, that is what we perceive as energy and therefor the essence of energy is how systems undergo change. If you look into the Wolfram Model of Physics, there is a mechanism of this notion that energy is "the density of things undergoing change." It is the true model of physics, and i would say that anyone that seriously studies these particular fields about complex systems will reach the same conclusions. Hoffman's ideas, are also in correspondence with the wolfram model (the network conscious agents in Hoffman, is the same as Wolfram's Atoms of space, which when following rules, create a complex hypergraph...except atoms of space are computational constructs, where as hoffmans are well...not...and thats where they differ.) The reason to take the Wolfram model more seriously is because it answers MORE questions that those other theories, as correspondent as they are, don't answer some bigger more pressing questions, that the wolfram model does. Cheers,

@Carlos.Explains Год назад

Wow, this is fascinating. I need to do a much deeper dive on Wolfram's model. I got a few chapters into A New Kind of Science but quickly hit the limits of my understanding.

@NightmareCourtPictures Год назад

@@Carlos.Explains I would suggest rather than reading the book (even though that's cool, its just a really thick book) is to watch the 13/14 part series on RU-vid he made on it instead...where he reads the book but also explains his thoughts on the concepts while doing so, and it's easier to digest the contents. Quick summary of the book; The first part of the book, is that he ran a computer experiments, looking at programs in a general sense, to see what they do and analyzed their behavior. He then runs proofs of exhaustion, on these programs, rather than just observing what any particular program does, he would run entire classes of them and then analyze their behaviors overall. Most notably, that incredibly simple programs, tend to exhibit extremely complex behavior. The 2nd/3rd part of the book is making sense of those experiments...and at the final chapter, deriving from that analysis, his principle of computational equivalence, which is the idea that nearly all systems following rules, reach a maximal level of complexity...that ceiling for complexity being Turing universality: The ability to compute all computable functions. When such a thing is true, it has major implications on...just about everything imaginable. The most striking thing about the principle is that it is an equivalence statement between all systems (similiar in scope to how space and time were unified by, showing an equivalence to all frames of reference; Lorentz transformation under the notion that light is constant) In the case of computational equivalence, what is constant in this picture, is the ability for systems to follow rules...that in essence any system can follow any set of rules and be arbitrarily complex...and this means that all systems are dipping into a sort "shared rule space" a single construct which is the Ruliad concept Wolfram talks about : The total rule space of all turing machines. That is just the beginning of the huge ramifications for what such a thing means...and there's so much i could say, but i think giving it the time to learn and understand it will hit you as it did me, that all the concepts people have here (levin, Tegmark, Markopoulou, Gates, Susskind, Nima-Arkani, Thooft, Hoffman, and this fellow you interviewed) are all touching on aspects of the same thing, and they all converge with the wolfram model imo. Cheers,

@Carlos.Explains Год назад

@@NightmareCourtPictures Thanks for the summary! I gotta check out that series asap...

@thewaythingsare8158 Год назад

I had a very interesting episode with my sister who claimed the my bag was speaking to her. But when I explained to her that the bag did not have vocal cords or lungs in order to activate those cords, in order that the air would vibrate and those vibrations enter her ear and vibrate her ear drums etc etc, the voice from my bag stopped. Its as if I managed to put back some kind of signpost in my sister's brain that said the voice was internally generated and could not be generated by a bag with no physical capacity to speak.

@xonack Год назад

interesting point about the luck mindset facilitating actions that will confirm that mental frame - law of attraction playing out psycho-physically? incredible content Carlos - thank you 🙌

@Carlos.Explains Год назад

Thank you, xonack! I think there's a likely link between positive mindsets and results.

@yoloswaginator 10 месяцев назад

„Playing out“ sounds like it explains a specific circumstance, but attending to avenues minimizing prediction error given our priors fully explains away this supernaturalistic priciple. Knowing that being overly optimistic is a great aid to actually succeeding is well established in many sports and professions, as well aw other aspects of life, e.g., learned helplessness is just as much an example of self evidencing.

@mrc4352 Год назад

Hello, thx for ur effort to explain things.......the videos i found and watched seem to apply to me... but im no scientist at all... it just happend and the more i find in the Internet there are some strange connections in fields/subjects i have actually nothing to do with me.....

@mri8517 Год назад

Carlos putting the cube there with the dude seeking guidance is a good picture.

@WalterSamuels Год назад

I think I can answer your lotto question, which was a great one and got me thinking. Like you said, perhaps it’s the word “surprise” throwing things off and I agree, it’s not a great word to describe the principle and the cause of the confusion. If you had 1 in a billion chance of winning the lotto, you’d be very surprised if you won, but the result of winning still has a fixed utility relative to the amount won. If you had a 1 in 20 chance of winning the lotto, you’d still be surprised (less surprised), but the utility is the same if the winning was the same as in the previous scenario. So the surprise in this scenario is purely just an emotional reaction to the result, but not relevant to the outcome. But if we think of surprise in the prediction error rate reduction context, our goal would be to enter the lotto with better odds of winning (the 1 in 20 case), thus we’re “minimizing surprise (reducing prediction-error)”.

@mootytootyfrooty Год назад

Great interview, really touched all the really important areas around Karl's work in a more down to earth way. And I'm really glad you keep asking about syncretisms with others' work, please keep on that track. Might be interesting to try to get in touch with some down on the ground hardware or software engineers in the QIT space, or maybe industrial adaptive systems. It gets super weird super quick but trying to ground the conversation together with discussions like this could be a really interesting challenge since they're for the most part dealing in radically new applications of all this math on actual atoms and photons, or even giant chains of molecules and larger fields (simulated or actual). They are constantly the ones giving the necessary feedback and testing for work like at Karl's lab so there's a different nuance in their understanding.

@Carlos.Explains Год назад

Amazing recommendation, thank you Joshua! Wasn't even on my radar before, I will definitely look into this.

@mootytootyfrooty Год назад

@@Carlos.Explains automod removed my linked comment but I'd love to see Edward Grant or Andrew Hallam or anyone on their Hierarchical Quantum Classifiers papers. Obscure stuff, seen basically nothing about it. The QIT world is a bit humdrum with cryptography or satellite etc but the computational guys like thkse two I mentioned or perhaps some of the people doing sub nanometer process control are definitely on an interesting trip. So much quantum tech is every day stuff without really realizing, e.g. flash memory.

@Carlos.Explains Год назад

@@mootytootyfrooty Thanks for the tip! And strange about automod... I thought I had that turned off for all my videos.

@WalterSamuels Год назад

This was a great discussion, incredible as per usual. You have better content than the likes of Joe Rogan and Lex Fridman as far as I'm concerned. I do wonder if we're missing something with this principle though. On the surface it doesn't seem to account for the behavior of drug use to me. For example, I know that when I drink alcohol it "relaxes" me, it doesn't lead to any new insights (maximizing information or exploration), I just feel a state of comfort or bliss. But I also know that my hangover the next day is going to be terrible. Regardless, I do it anyway. To me, this feels like a trade-off in short term vs long term reward. In the short term it's a reward, in the long term it's a punishment. Despite knowing that the long term punishment outweighs the short term reward, I still choose to drink alcohol. What is the explanation of this within the framework of the free energy principle? I admit I may just be failing to draw an obvious connection, but if the principle is correct it should be able to describe all forms of behavior to a degree.

@sunnyinvladivostok 11 месяцев назад

Interesting comment you have, interesting to think about. I'm just a youtube commenter, I know little about drugs or the free energy principle, but I do wonder if there is a separation between our observations and the implementation of the principle.. maybe better said that the FEP loses its coherence when navigating temporal space. i.e. me not saving for retirement right now will create an immense amount of uncertainty for me in the distant future, but the feedback loop is too long for me to make any internal correlation. idk. Another thought is that this is why drugs are seen as drugs, because they short-circuit these mechanisms our biology has come up with.

@dfearo Год назад

Thanks!

@Carlos.Explains Год назад

Thank you so much!!

@ajsindri2 Год назад

OH HELL YES!

@MrPDTaylor Год назад

Oh heck indeed

@ChristianSt97 Год назад

great!!!! btw Carlos I wanted to ask you if you are going to interview Bernardo Kastrup one day.

@Carlos.Explains Год назад

Hey Christian, thanks for the comment! I inquired with him but unfortunately he declined. I need to get the subscriber count up. :)

@ChristianSt97 Год назад

@@Carlos.Explains will surely happen some day soon

@Carlos.Explains Год назад

@@ChristianSt97 Thanks man! I so appreciate your support... got a lot more coming!

@real_pattern Год назад

@@Carlos.Explains he's appeared on several channels with significantly less subscribers, so it's probably not about that, almost certainly not just about that.

@Carlos.Explains Год назад

@@real_pattern I think he's still doing interviews with podcasts he previously appeared on, regardless of subscriber count. Or maybe he's terrified of my hard hitting style. 😂

@davidrandell2224 11 месяцев назад

The expanding electrons do it all. No energy, charge, photons, waves, spin, fields, potential, quantum,quarks, space, time, space- time, etc. All Standard Theory/Model was replaced by Expansion Theory in 2002. “The Final Theory: Rethinking Our Scientific Legacy “, Mark McCutcheon for proper physics.

@CipherOne 10 месяцев назад

Please use a de-esser on your audio.

@jbperez808 2 месяца назад

Why is it unfalsifiable? Isn't it enough to show a case where the equations cannot be found to work/apply even in a situation where a "markov blanket" can be said to exist to falsify it, at least for that particular case? (eg that it cannot be said to be universally applicable?)

@Carlos.Explains 2 месяца назад

I think because it's more of a framework than a theory per se. I'm not sure there's a case with a markov blanket that could falsify it...

@TimeDeltaNova Год назад

55:55 I don’t fully understand why information gain and surprise minimization aren’t balancing forces. It seems to me that if they were both operating in competition with each other, this would lead to the desired outcomes of avoiding death but also not sitting in a dark room forever.

@Carlos.Explains Год назад

I haven't fully grokked this yet. I think a system tries to maximize information gain while minimizing surprise.

@waylonbarrett3456 Год назад

I've built models computing something very much like active inference and, at least in my particular implementation, these two forces do work against each other to produce fascinating outcomes. I wouldn't say they balance each other. And, you wouldn't want to actually reach equilibrium anyhow. Equilibrium-adjacent phenomena are where all the interesting stuff happens. In my models, the drive for increasing information gain is further abstracted as "empowerment" (information theory concept) in the context of modeling affordances in the agent's environment.

@Carlos.Explains Год назад

@@waylonbarrett3456 Thanks! I recently learned about empowerment in Bobby Azarian's book, The Romance of Reality (interviewed him a few days ago!). I definitely want to dive into this concept in the future.

@real_pattern Год назад

the folk psychological formulation of the free energy principle: f**king around and finding out.

@thebagtalksulisten Год назад

Bayesian mechanics is absurd

@Carlos.Explains Год назад

I just consulted my priors and agree.

@jbperez808 2 месяца назад

1:07:00 Proof that God is in hell! 🤣🤣🤣🤣🤣

@Carlos.Explains 2 месяца назад

A lot of interesting implications arise from full knowledge...

@davidkemp3154 Год назад

The paradox makes most money in the world is that supposing a person is smarter then that he is also more productive & truthful. As humans multitask more unfortunately America loses its grip on growing the honest man. Focusing on being smart & more truthful is a paradox that may not be attainable like trying to accelerate to avoid being eaten by a black hole. Impossible unless you are quantum entangled radiation. Nevertheless a message that can convey across borders & generations is the "pursuit of happiness" as America's godgiven founders put it long ago. Lawyers be damned!