Вставка
- Опубліковано 14 чер 2024
- This week Dr. Tim Scarfe, Dr. Keith Duggar and Connor Leahy chat with Prof. Karl Friston. Professor Friston is a British neuroscientist at University College London and an authority on brain imaging. In 2016 he was ranked the most influential neuroscientist on Semantic Scholar. His main contribution to theoretical neurobiology is the variational Free energy principle, also known as active inference in the Bayesian brain. The FEP is a formal statement that the existential imperative for any system which survives in the changing world can be cast as an inference problem. Bayesian Brain Hypothesis states that the brain is confronted with ambiguous sensory evidence, which it interprets by making inferences about the hidden states which caused the sensory data. So is the brain an inference engine? The key concept separating Friston's idea from traditional stochastic reinforcement learning methods and even Bayesian reinforcement learning is moving away from goal-directed optimisation.
Remember to subscribe! Enjoy the show!
00:00:00 Show teaser intro
00:16:24 Main formalism for FEP
00:28:29 Path Integral
00:30:52 How did we feel talking to friston?
00:34:06 Skit - on cultures
00:36:02 Friston joins
00:36:33 Main show introduction
00:40:51 Is prediction all it takes for intelligence?
00:48:21 balancing accuracy with flexibility
00:57:36 belief-free vs belief-based; beliefs are crucial
01:04:53 Fuzzy Markov Blankets and Wandering Sets
01:12:37 The Free Energy Principle conforms to itself
01:14:50 useful false beliefs
01:19:14 complexity minimization is the heart of free energy
01:23:25 An Alpha to tip the scales? Absoute not! Absolutely yes!
01:28:47 FEP applied to brain anatomy
01:36:28 Are there multiple non-FEP forms in the brain?
01:43:11 a positive conneciton to backpropagation
01:47:12 The FEP does not explain the origin of FEP systems
01:49:32 Post-show banter
www.fil.ion.ucl.ac.uk/~karl/
#machinelearning
Pod version -- anchor.fm/machinelearningstre...
![#67 Prof. KARL FRISTON 2.0 [Unplugged]](http://i.ytimg.com/vi/xKQ-F2-o8uM/mqdefault.jpg)
![#67 Prof. KARL FRISTON 2.0 [Unplugged]](/img/tr.png)
This is a lot more useful than lex Friedman. Fascinating talk with Friston, really appreciate the background work you put into it.
Great podcast, but the music is really, really distracting
Skip to 36.33 the entire interview is there unadulterated
Geawd. No kidding. So. Much. Noise.
Trying hard to get past it. So hungry for this content and the overall production quality is otherwise exceptional.
No music please!
You guys are amazing! Your preparation really shows. What's even more amazing is the artificial life that is the UA-cam. Months ago it introduced me to Hameroff's idea of quantum effects in microtubules for photosynthesis. Following up yesterday morning I searched for quantum mechanics/least action principle and UA-cam supplied Feynman's PhD thesis anticipating his work on QED. Yesterday afternoon it offered me, unbidden, Friston, who mentions Feynman, path integrals, and what seems to be the Holy Grail, Free Energy Principle. By manipulating a few electrons here and there UA-cam seems to be a conduit for effecting communication among the collective consciousness that Dean Radin has been monitoring. True synchronicity with a concrete mechanism exposed. Amazing!
Really great job! What I love about Prof. Friston is that regardless of how great and accomplished he is, he always stays in touch with the community. Really great!
Such a powerful line. "Objective functions should about belief and consequences of belief states, not states of the world."
So having watched and revisited a few segments-
There is so much to say here, these subjects are heady and far-reaching. But first of all thank you for bringing Prof. Friston as a guest and providing excellent leap-off questions. I think the main thing I find is that *how cool is it* that we are at a point in time- scientific groundwork and technology wise -that we can even entertain seriously the idea of applying first-principles style modeling and mathematical formalism to this subject matter and the deepest questions we've ever thought to ask?? I would wager most of the audience here has grown up with these questions and some less-or-more rigorous version of these thoughts swimming about their heads. And, seemingly, we are at the point where it is practical to invent or experiment to harness and improve a formulation that combines theoretical physics and intelligent agency, without it being nonsense on a cartoon villain's secret laboratory blackboard. :D
I also find it very encouraging that there seems to be so much usefulness that can come from investigating these ideas. Even if we never arrive via this pathway at some sort of truly universal explanation of the dynamics of embodied intelligence, all the math and contemplation that goes into it will bear fruit in "workaday" applications that can benefit from, say, markov blanket modeling or whatnot. Great stuff.
I really enjoyed the edited summary at the beginning of this one.
The background Vivaldi four season just gives such a nice expectation to the talk
Regarding the question about learning phonemes from the radio vs. from speech in social interaction, I believe the difference is well explained in Stanislas Dehaene’s “How we Learn.” The difference between these two situations is shared attention. In a social context a child is going to attend to the noise produced by adults and see it as important because it will recognize them as sources for the noise. In the radio situation, unless the child already speaks and can recognize the noise as language, it will not be attending to it in the same way. Dehaene describes a slightly different situation, learning an association between an object and a word from an adult vs. a loudspeaker, but I think the general principle is the same. Young children are hardwired to attend to what adults do (gaze direction detection, facial recognition, etc.) and attention, as we all know, is all you need.
Why why do you add the music? Do you know how difficult you make it for us the aveverage intelligence curious person to listen absorb and understand?!
Please release on Spotify a music free version.
Wonderful. Thanks for creating and posting.
I enjoy the chat. Very informative, NEW WAY TO SEE THINGS THAT ARE HIDDEN,.
More subscribers to you, really such nice work. UA-cam algo must be stupid not suggesting your content. But this will follow and then you will have a peak
PLEASE GET RID OF THE BACKGROUND MUSIC. It isn't necessary, and is distracting.
This is Great information! Thank You so much!!!
Nice one :) love it
Very interresting topic ,but i find the background music in the first part very distracting..sometimes even overwhelming what is being said!..the other part is totally fine
we are noobs at editing, it's a process of evolution :)
I have am a Physicist. Coincidence or not I am working independently for about a year on the problems of LLMs, The approach I am developing is actually very similar to Dr Karl. Good to know there’s already a framework for that. I am well versed in numerical analysis, multiple scale methods and programming. My academic work was basically on objects that I think are very suitable to FEM. Solitons in PT symmetrical systems. Not any system though. I found a particular type of system in a particular kind of bifurcation that behaves like being alive. Like, instead of carefully prepared, they actually form themselves from very weak initial conditions. Also pretty hard perturbations very quickly dissipate from these solitons. They cure themselves. I really would love to collaborate with someone on this topic. My papers are easy to find. F. C. Moreira SHG solitons on scholar will lead to results.
14:24 "one of the few principles that conforms to itself" -- just a side curiosity: in words, these are known as autologies. 'Pentasyllabic' or 'noun', or 'repugnant'
Thanks DavenH!
@ 1:06:10 See Arizona State University 1997 Masters Thesis, "Reinforcement Learning: Experiments With State Classifiers for Controlling an Inverted Pendulum", Figures 5.9, page 74, and Figure 4.1, page 53, where the fuzzy inference rules to model the physics (torque direction) are generated automatically.
Thanks
Music at 18 mins is distracting as hell when you're a beginner trying to parse what the hell you guys are talking about...but thank you for the episode though, amazing as always
Great chat. Thanks for the time you focused on a sensible exchange of thoughts. Also: "Skit on cultures" = "techbro/jerk bs".
"Objective functions should about belief and consequences of belief states, not states of the world." WOW
So if something with a Markov blanket system exists over time, it must resist entropy by gathering evidence of its own existence, minimising free energy or surprise, via a change in its modelling or actions. The first line of thinking makes sense, if something exists over time and is consistently able to resist entropy, it must have a strategy to do so whether consciously or not.
Whether that strategy is gathering evidence of its own existence, I'm not sure.
For example, I'm not trying to confirm my own existence by any means, I'm looking to understand it, even if understanding it moves me towards an increasing accumulation of evidence that disproves my existence. I'm not actively seeking the blue pill if you will, I seek the red as it moves me closer to the truth. This may be my strategy for resisting entropy, by better modelling the world, others and my self, but the aim isn't reduction of uncertainty, it's growth. Growth in consciousness. Growth in level of awareness of being, in relation to understanding of self, other (people/species) and whole (nature/universe/God).
Perhaps this reduces down to a systems ability to process information at the various Markov blankets in which it exists. Or awareness of that within its blanket, outside its blanket & the interface between the two. The quantity and quality of its data set, the ability to process, model & synthesise that information, then translate those learnings in to revisions of its modelling or actions, in a real time continuous learning loop.
Minimising uncertainty might be a key survival strategy, but is mere survival the highest motivating force in the universe?
"in some realist sense, we don't think that the photon travels every one of those possible paths, but it's 'as if' it did."
I'd like to challenge this. The story is a bit more subtle than that. Let's take the simplest case of a Mach-Zehnder interferometer which is the same idea but with only two paths (two inputs into a first beamsplitter, two mirrors with tunable phase recombining the outputs into a second beamsplitter and we look at the two final outputs). The state of the photon in the interferometer can be represented using the Bloch sphere: the North Pole represents travelling in one arm of the interferometer and the South Pole is travelling in the other arm. The states on the equator are equal superpositions of travelling in either arm with an adjustable phase (the longitude). We can show that the photon _actually_ goes both ways because if it didn't it wouldn't feel the phase difference (because there would be no "equator"). We actually detect the fact that it feels the phase difference by looking at the statistics with which it emerges from the final outputs. Closely related to this, there are some paradoxical gedankenexperiments (a few even tried out for realz usually in quantum optical experiments) like the Elizur-Vaidman bomb tester, and many things in the field of interaction-free measurements: there is a _very real_ sense in which we can say that the photon _actually_ went through all of the paths.
I am attracted by the aspiration of mathematical psychology. For me it is not just explaining what things are but also what the thoughts are in a relatively indisputable manner. I.e. in mathematics, compared to other perspectives, like from personal experience, subjective minds, which are 'various'. However, it is a difficult conversation for me to follow.
- the headline is about "Free Energy Principle", but where is the "Energy"? All I heard was things about information, bayesian method etc., but nothing literally on "Energy".
- reading wiki page about "Free Energy Principle" seems opening more questions than closing. To understand "Free Energy", one has to know the 2nd thermodynamics and "minimum energy principle" and its relationship with information theory.
- "minimum energy principle" explains the isolated system with constant entropy and closed system with constant energy. I think energy is easier for me to dig first than entropy. But what is energy? and what means having constant energy of a system?
- It is a long way to go, I guess.
Believing is easy, understanding is hard.
For the origin of life and intelligent systems, I recommend Jeremy England’s work. His research sheds light on how life and intelligence (predictive computation) can spontaneously occur as a consequence of the second law of thermodynamics and the fluctuation theorem.
@2:50 the Feynman SOH does not average over paths. It uses interference, which is not an averaging process. It's better to think of the Path integral formalism as a generalized Fourier transform from paths to amplitudes. It's utterly different to statistical mechanics. However, there is a clear analogy between QM and SM, you just replace probabilities by geometric amplitudes, then the Boltzmann partition function becomes the quantum partition function or "Quantropy" as coined by Baez and Pollard. The problem is the geometric amplitudes make QM utterly different to SM.
The music may be very good, but in this use it is noise
Free energy principle reminds me of my hypothesis: "Research Gate/Why the purpose of the species is to create Artificial General Intelligence"
OMEEGAD !!! Thourf !
Isn't it possible that a hurricane does have a Markov blanket, but the dynamics are changing so quickly it is beyond our sensory perception to make sense of the boundaries?
omg, Third!!! 🙌🙌
haha :D
As a doctor, I am quite interested in the free energy principle but feeling like my math and physics knowledge is not enough to fully understand that. Do you have any book recommendations for my situation? I really want to dig into this.
Hey! To understand the maths which Friston uses, read this book www.amazon.co.uk/Data-Analysis-Bayesian-Devinderjit-Sivia/dp/0198568320
The state space of understanding is in both feynman diagram, as well as model physics and math. The diagram gets u there much faster, though
Well I am a Physicist, but more like on the mathematical methods an computational side. I am really open to collaborating as I am not on academia anymore. I do research and development by myself.
what makes you think tank music helps understanding let alone comprehension?
Summer + Friston + MLST never liked so fast 😆
How does addiction (drugs or of other sort) and OCD fit into this?
Now like I'm a retriever :)
bring back Conor
We hope he will come back and do some shows with us soon too!
Second!!
Have you ever heard the phrase: “He has southern efficiency and northern charm” 🙂
Is that kind of like "A face for radio [and a voice for newspaper]"?
@@DavenH excellent example, I’m curious to know if GPT-3 can grok that one ☝️
music sounds great at the intro for a brief time but soon disturbs the flow a bit. try some lofi beats maybe haha
I disagree. Love the music. It makes me feel excited about what I’m about to learn ¯\_(ツ)_/¯
@@_tnk_ I get that some might like it but I am not able to listen to the program because of the music. Maybe it’s a matter of age, but I cannot express how distracting the music is for me. Glad you like it, but I don’t think I am alone in not being able to tolerate it.
Larger entropy and lower free energy?
52:00
57:00 is beautiful
I watch these at 1.75x and the music is a God damn disaster. Please NEVER EVER do this again.
Lol! 😂😂 As you might have noticed we are using it far more sparingly these days
Do you need an alpha? - "Absolutely not." Great, I've been trying to eliminate it for decades, now Friston has a way of getting rid of it w/o cross validation, etc. "Or absolutely yes." Dang it.
Your young age is showing. Describing Friston as "old school" professor an "eccentric" who sits in a Chesterfield chair with spring coming out. Are you kidding?! Friston is a genius and is inventive and creative in his thinking as any person anywhere.
Your lack of reasoning is showing, we also think that Friston is a genius, is inventive and creative in his thinking. Making a light-hearted joke about his eccentric style only serves to reinforce this. Keith didn't like the joke at the time and we probably would have thought better of it if we could go back
While I appreciate the bureaucracy of machine learning for giving us a cooperative endeavor and giving us much to debate and agree upon, I prefer to think that an intelligent machine will be created so to speak in the garage by one individual genius. Examples are Mark Zuckerberg inventing Facebook, Steve Jobs pioneering Apple, Nikola Tesla inventing alternating current, and Henry Ford introducing the Model T. That being said, could I be the one to go down in history as inventing the intelligent machine? I have my theories, and admit I have much work to do, but I'm working diligently on the problem as we speak. In other words, there are just some things where many cooks spoil the broth. Could you be the one Neo that finds the key to intelligent computers? I strongly suggest that you try your best, but remember history. History favors those who try and put ideas into action, not those who talk, talk, talk. I see in the future putting our best neural nets into a Boston Dynamics robot and taking over all the hard, dirty jobs that no human wants or has the capability to do. Whether academia will find the most efficient thinking machine or the lone genius will, time will tell.
What is the point of understanding your brain if you can’t use that to understand the nature of reality? Our brains are comparators, constrained to dualities like mind/body, love/hate, … - so are there limits to our understanding of reality? Our thoughts are fundamentally no different than the bony skull surrounding our brain - just different complexitites of energy patterns. Space and time cannot be dissected apart. So It occured to me during this discussion that perhaps one can conceive of reality is a Markov blanket of energy fluctuations. Particles themselves are fields of energy in (if they are stable) steady-state equilibrium, but able to exchange energy via fields with other particles. The point is, our concept of space as 3-D or 4-D or whatever is naive. It’s non-dimensional; it only seems dimensional because that’s how our brain
models it. Of course we think of those cute little markov diagrams as 2 or even 3 dimensional, but what about the energy fields and interactions they represent? If energy doesn’t have a dimension, how can particles? Because that’s how we measure them. We view a particle as a point because we can’t discriminate energy fluctuations within it. A blurry point (a wavefunction) at that. All we see are 2-dimensional images, but we construct 3-D by layering those 2-D images over time. Without time, space has no meaning. And time is the essence of Markov interactions. QED. Ha ha!
A fantastic program rendered unlistenable by intrusive, distracting, and completely unnecessary background music. I just so wanted to dig into this program, but was not able to listen to it. Why would you do this? The music adds nothing and detracts from the utility of this otherwise excellent production. Do you think you could post a version without the music?
Hi Steven, I'm sorry to hear you didn't enjoy the (long) "intro". Might I suggest you skip forward to time code 36:33 ua-cam.com/video/KkR24ieh5Ow/v-deo.html where the main show begins? There you will in fact be able to listen to the entire content recorded with Prof. Friston without music. I hope that helps and apologies for the delay in responding.
First 😅
I saw one person who knew what he was talking about and three people who pretended to know what he was taking about, but were completely snowed.
I think I see what's going on here. It's similar to the 90s, where the linguists tried to take over all epistemology, and almost convinced the academic world that they were correct. This is now the chemist's run at it. What gives me pause is that they can't explain a single concept without dependence on jargon spruced up with mathematical equations.
This this just gobbledygook. Please prove me wrong.