This man builds intelligent machines

Bert de Vries is Professor in the Signal Processing Systems group at Eindhoven University. His research focuses on the development of intelligent autonomous agents that learn from in-situ interactions with their environment. His research draws inspiration from diverse fields including computational neuroscience, Bayesian machine learning, Active Inference and signal processing.
Watch behind the scenes with Bert on Patreon: / bert-de-vries-93230722
/ discord
/ mlstreettalk
Bert believes that development of signal processing systems will in the future be largely automated by autonomously operating agents that learn purposeful from situated environmental interactions.
Bert received his M.Sc. (1986) and Ph.D. (1991) degrees in Electrical Engineering from Eindhoven University of Technology (TU/e) and the University of Florida, respectively. From 1992 to 1999, he worked as a research scientist at Sarnoff Research Center in Princeton (NJ, USA). Since 1999, he has been employed in the hearing aids industry, both in engineering and managerial positions. De Vries was appointed part-time professor in the Signal Processing Systems Group at TU/e in 2012.
Pod version: podcasters.spotify.com/pod/sh...
Contact:
/ bertdv0
www.tue.nl/en/research/resear...
www.verses.ai/about-us
Panel: Dr. Tim Scarfe / Dr. Keith Duggar
TOC:
[00:00:00] Principle of Least Action
[00:05:10] Patreon Teaser
[00:05:46] On Friston
[00:07:34] Capm Peterson (VERSES)
[00:08:20] Variational Methods
[00:16:13] Dan Mapes (VERSES)
[00:17:12] Engineering with Active Inference
[00:20:23] Jason Fox (VERSES)
[00:20:51] Riddhi Jain Pitliya
[00:21:49] Hearing Aids as Adaptive Agents
[00:33:38] Steven Swanson (VERSES)
[00:35:46] Main Interview Kick Off, Engineering and Active Inference
[00:43:35] Actor / Streaming / Message Passing
[00:56:21] Do Agents Lose Flexibility with Maturity?
[01:00:50] Language Compression
[01:04:37] Marginalisation to Abstraction
[01:12:45] Online Structural Learning
[01:18:40] Efficiency in Active Inference
[01:26:25] SEs become Neuroscientists
[01:35:11] Building an Automated Engineer
[01:38:58] Robustness and Design vs Grow
[01:42:38] RXInfer
[01:51:12] Resistance to Active Inference?
[01:57:39] Diffusion of Responsibility in a System
[02:10:33] Chauvinism in "Understanding"
[02:20:08] On Becoming a Bayesian
Refs:
RXInfer
biaslab.github.io/rxinfer-web...
Prof. Ariel Caticha
www.albany.edu/physics/facult...
Pattern recognition and machine learning (Bishop)
www.microsoft.com/en-us/resea...
Data Analysis: A Bayesian Tutorial (Sivia)
www.amazon.co.uk/Data-Analysi...
Probability Theory: The Logic of Science (E. T. Jaynes)
www.amazon.co.uk/Probability-...
#activeinference #artificialintelligence

[@muhokutan4772]

This is probably one of the best MLST episodes, each new episodes feels like a long awaited reunion with a loved one, this work is immaculate and invaluable!

[@diga4696]

You are absolutely correct! It feels there is a sentient AI producing MLST episodes just for me, based on its understanding of me. This is better than Netflix! I love this episode, the langrangian has been a fascination of mine for almost 15 years.

[@madmanzila]

felt the same ..these are some really important conversations ...

[@ArchonExMachina]

What he is describing makes so much sense from a practical technical perspective. I think this approach will be the next big thing.

[@paxdriver]

I think we may be overdue for an in house episode where Tim and Keith can hash out some ideas they've gathered from all these great podcasts. An annual spitball would be a great tradition to consider, perhaps. You can't possibly talk to all these geniuses and work all year on the cutting edge without building up a whole 3hrs of blended ideas from the last year. Keith's idea of individuals as instantiations or trials as part of the composite whole model of natural selection is especially illuminating in just this way (~ 1:00:00) relating to active inference and concurrency queues after that. It's just so insightful and helpful to help the mind stay jelly when we all tend to focus on the specific goal and deadline. Like Burt reading other papers with relevence in mind, these kinds of talks make everyone in the field better at thinking about everything else they work on and study. Even the philosophy stuff that doesn't come up in topical conversation, I bet you both have a tonne of ideas like that I bet viewers would love to hear. I know you both look up to all of your guests, but I think you maybe sometimes neglect the novelties of your contemplations sometime. It'd be awesome to hear more about applying active inference to async software engineering, like chunks, thread pooling, or other variations of attention and transformers. I bet you guys got a ton of wickedly interesting discussions off camera over a pint.

[@roseproctor3177]

I so agree! imagine a livestream of engineering something 😍

[@MachineLearningStreetTalk]

I can hear the sound, it will be there for everyone when the video processes to HD I think

[@asimuddin3222]

Now it is good. Much appreciated

[@betel1345]

I love your presentation. So imaginative and clear, interweaving your guest, the venerable Karl and your elaborations. Rich. Excellent. Thanks

[@pennyjohnston8526]

Feels like we've moved into the implementation phase for FEP! Liked the multi modal narrative, back drop, delivery, pace, hit a new level of engagement and experience ! Tim's virtual library in the discord with click throughs/reviews? Thank you MLST for all the work - yet again !

[@MWileY-nj1yb]

I deeply appreciate you guys and the superb work you do. Many thanks and much love.

[@mus3equal]

This was phenomenal thank you all!

[@youknowwhatlol6628]

Thanks! Without music, it's not distracting. Very interesting, thank you so much!!!!

[@Johnmoe_]

man this podcast has insane quality HOLY

[@sharkbaitquinnbarbossa3162]

Great Talk! Very informative and accessable.

[@ashred9665]

Very dense topic, very high quality.

[@petersuvara]

There's one thing that flies in the face of the optimisation problem and Machine Learning in General. The idea of creativity and artistic expression, since it's not bound to optimisation, it's an expression of the state of things in all of it's varierty. Like the left and right brain, like the idea of order and chaos. There's an intrinsic duality that machine learning needs to connect with in order to become something more than an optimisation problem to be solved.

[@missh1774]

57:40 the agent should be able to say. "No thank you, not right now" and it should be able to deactivate or put a sleep mode on the running connection. Wonderful interview conversation. Thank you!

[@todprog]

Tim, the more I listen about FEP/AIF from your impressively crafted cinematic videos, and the more I study FEP/AIF literature, the more evidence and more researchers confirm the principles of the "Theory of Universe and Mind". FEP/AIF is a more technical and operationalized version/line of research and an elaboration of the core principles from that earlier interdisciplinary body of work, first published 2001-2004. TOUM was taught as the ultimate lecture during the world's first university course in Artificial General Intelligence, presented in 2010 and 2011 at the university of Plovdiv. An epoch ahead, but barely recognized. Weirdly that theory was invented and the courses happened a few kilometers away from two emblematic "Markov blankets": the village of Markovo, a walking distance away from Plovdiv, known as "Plovdiv's Beverly Hills", and one of the famous seven hills in Plovdiv, called "Markovo Tepe", now converted to "Markovo Tepe Mall".

[@muhokutan4772]

Thanks!

[@sapienspace8814]

@ 1:16:00 that is a great example of learning to ride a bike. There are at least three key control action output signals learned, one the handle bars, two, the pedalling, and a third is leaning the body relative to the bike as an inverted pendulum. Our bodies have a large number of action outputs, but the state space can adapt to focus attention (such as by K-means clustering with a cost function) on the small set of actions necessary to control the bike. Interestingly, the minimal set of actions is one for a very skilled rider, that is body position relative to the bike can be good enough to balance the inverted pendulum (though it takes great skill!). This is how the model can be updated on the state space to optimize the control function (e.g. K-means clustering on actions of highest interest/value with minimum cost, minimum necessary energy). This is likely how babies learn to crawl, then walk, then run. They adapt by learning to focus attention on optimal (minimum energy) action output sequences that produce rewards. It is like adapting and focusing a telescope onto the state space.. Basically saving an updated model weights after a certain number of runs, like while dreaming over night, a pruning process, or like defragmenting the "hard drive" of the mind. Or like identifying the position of stars in a constellation and writing down their coordinates for observation tomorrow night on paper as they are interesting.

[@sapienspace8814]

@ 17:28 Using Reinforcement Learning (RL) with Fuzzy Logic (self incriminating terminology), inference rules can be derived automatically and be optimized via a cost function of the state space (this is a different cost function than what is normally used in RL to cluster and focus the state space into regions of interest). I did not know he was in Holland, I was there earlier this month, wish I could have met to discuss this, even though I am not likely of much significance. Very interesting discussion, thank you for sharing!

[@-mwolf]

"the ultimate gentlemen" just moved into my vocab

[@v-ba]

Great talk, thank you

[@mattgosden]

This was the most informative and useful of the Active Inference series. More practical and tangible ... speaking with an engineering hat on

[@MachineLearningStreetTalk]

Thanks Matt! We did make a conscious decision from the beginning to steer away from engineering content so this is a bit of a treat 😃

[@d.lav.2198]

Putting the FEP alongside the Principle of Least Action really turned a few cogs in my brain.

[@fixitorforgetit]

3 books and a paper, thanks for the recommendations!

[@maddonotcare]

Sound is good👍

[@sapienspace8814]

@ 1:34:20 a way to do that is to randomize the initial sparse state space node positions. I suspect biological neural synapses physically move themselves to nodes of interest over time (kind of like how Bert described lighting traverseses minimum impedance paths as described to dissipate the high voltage capacitive charge from potential to kinetic energy). This is also entropic convergence described by Klopf in his book "The Hedonistic Neuron". The neurological system seeks the lowest energy state (entropy). What is really fascinating to me is how a Paramecium can mate, move and eat without neurons. Hameroff points out it may be via microtubules (these also guide the separation of chromosomes during cellular division). There is very interesting work by Hameroff (University of Arizona) on microtubules (may operate at infrared frequency).

[@exhibitD79]

''mate, move and eat without neurons.'' Is that really what they are doing. Those words have concepts assocaited with thinking. Why are we calling the observations that though? Cellular level reactions can appear like one thing is being consumed by another thing. Is that really ''eating'' though? To me this is where the concept of words get's confusing. We have an ontological problem rather than something to explain about actions without neurons.

[@sapienspace8814]

@@exhibitD79 Very good point, it is all chemical reaction of inanimate molecules at nearly every level of context. To understand how to build a powerful "mind" is to understand its lowest fundamental level of physics that expresses itself in contextual, adaptive, motion, or harmonics. This is Klopf's hypothesis at the neuronal level up, though he did not look further down to lower contextual levels of the neuron. In Klopf's book, "The Hedonistic Neuron", he looked at higher contextual levels from neuron, to brain, to society. Microtubules go in the opposite contextual direction, to a context below the neuron. When the brain grows the neurons move. The synapses also move. Is it the microtubules generating and directing this motion? Microtubules control motion and structure of the cell, they have different expansion and contraction rates and control biological "motors" that move flagella. My mind had your same perception, however, earlier today I found a paper on how an amoeba is observed to have an "associative memory" based on some electrical stimuli experiment. I did not read the whole paper. The role of dynamic synaptic motion is crucial, it is effectively adaptive circuit wiring, can imagine a power grid that wires itself based on load demand by automatically increasing its capacity.

[@ArchonExMachina]

1:01:00 I'd like to have an episode dedicated to this discussion in language philosophy, it is quite intriquing. I'd like especially to hear Tim's take in depth, as it sounds like original thougth. It wouldn't have to be perfect, but just an entry in the discourse with your current views. Perhaps with a suitable guest(s). The notion of "a word as a conditioning force" especially interests me. Is this an active inference notion?

[@damienteney]

It would be nice to connect all these exotic (non-mainstream) ideas with concepts of machine learning that are much more established. Everything I head here sounds like things that are much more researched and have established names, like domain adaptation, unsupervised adaptation, anytime-computation, etc. And if they mainly aim at getting something to work (as said multiple times in the middle third of the episode), there's a real risk of falling back into the same local optimum of engineered solutions similar to what's been been done by others, even if they start from different fundamental principles.

[@palfers1]

Some fascinating perspectives here for this retired, physics-educated generalist engineer. It hit me like a ton of bricks that the Principle of Stationary ("Least") Action (PSA) cannot be derived from more fundamental principles. It would make a lot more physical sense if PSA, instead of being over Feynman's total paths, were expressible as a local measure that was calculated incrementally.

[@teleologist]

yo where da generative model come from? seems like magic.

[@EskiMoThor]

The principle of least action and biological systems made me wonder .. what about effort? Is there a measure of effort being tracked? Like ATP breaks down in our bodies, and adenosine causes tiredness, and in many cases the tiredness triggers growth/learning/optimization. Are analogous mechanisms used in active inference?

[@uncertaintyprincipal7119]

That thumbnail made me think Fury was looking trim before his up-and-coming fight with Uysk!

[@marcospiotto9755]

Nice talk. I am learning about kalman filters right now. Does anyone know some good implementation in python?

[@asdf8asdf8asdf8asdf]

Has anyone looked at Stephen Grossberg‘s approach to neural system development to see if there are any modularity or system interaction functions that would be helpful?

[@ehfik]

this podcast never fails to amaze.

[@FranAbenza]

Reinforcement learning with active inference. When? How? Analog computing?

[@KitcloudkickerJr]

I understand that my sentiment may not be shared by many, but that's perfectly fine. It's just not possible to have a perfectly deterministic system. We can try to understand the mechanics of a system by manipulating its nodes, but once we zoom out, the models we've created inevitably break down. The human mind, in particular, is a black box. Although we may know which regions of the brain light up during inference, we can never fully explain the thought process or reasoning behind someone's actions. There are 8 billion minds with 16 billion unique opinions and thought processes, making the mind similar to a neural network - a black box that works through convolution without a true mechanistic understanding. While we may lose abstraction while trying to gain mechanical understanding, I don't believe that aligning our understanding of intelligence with mechanics will work in practice, as it doesn't accurately reflect the true nature of intelligence.

[@leventov]

Agreed. I've expressed a very similar idea in a blog post "For alignment, we should simultaneously use multiple theories of cognition and value" (you can find it on the web).

[@KitcloudkickerJr]

@leventov ill be reading

[@exhibitD79]

Isn't that problem usually the differnce of missing information though. So it is ''possible'' it is just very extremly difficult.

[@Aandreus]

This is a visionary alchemist, and he is nested perfectly within an epic opening, well done! Today, we meet The Blair Wizard of phyical laws. If we do not percieve his project, indeed...we are doing the d@mn thing.

[@roseproctor3177]

I friggin love this podcast

[@steveshultz608]

Will there be a follow up video with additional interviews from your visit to the Verses offsite meeting?

[@MachineLearningStreetTalk]

Yes

[@steveshultz608]

@@MachineLearningStreetTalk looking forward to that…thanks!

[@sapienspace8814]

@ 1:38:30 I perceince it as we humans have built a scalable "Lego" of intelligence (and eventually the agents will even start designing its own intelligence Lego types if it is not already now). It seems this may actually be happening now before our eyes (it makes me wonder why we are alive when our recorded history is profoundly short, it makes me wonder if we are missing history, or we just so happen to be in a profound moment in human history). It is amazing that just four molecules, A, T, C, G to make DNA when placed into a structure pattern produce consciousness from inanimate molecules. From profound simplicity emerges extraordinary complexity, and consciousness.

[@timosalo5003]

1:28:50 ”Thunderbolt steers all things.” (Heraclitus)

[@wp9860]

I'm some 37 minutes into this video, but I just wanted to comment on the hearing aid problem. The question I have is how does the device receive its sensory information that is necessary to calculate hearing aid error? I envision that the hearing aid takes in audio signals from its environment and then sends, most probably, a different audio signal to the ear that the (corrupted) ear will process into an impression of the sound that is relatively the same as the impression a person of normal hearing would produce without the hearing aid. How is this perception by the hearing impaired person fed back to the hearing aid, allowing the hearing aid to calculate its modeling error? This happening while the hearing aid is continuously in use. The device being one "thing" with its own Markov blanket, and the person being the the hearing aid's environment of latent variables. Or, am I perceiving the problem all wrong?

[@pdsnk1]

Right. The 'agent' within the hearing aid needs to interact with the user to effectively tailor its function. Over time, this agent develops a model of the user's hearing preferences based on feedback received in various acoustic environments. There are several methods for this interaction between the agent and the user. For instance, consider a smartwatch that the user can discreetly tap if they are dissatisfied with the current hearing aid settings suggested by the agent. Additionally, there are advancements in integrating EEG with hearing aids. This technology can determine if the user is comfortable with the settings without requiring explicit feedback

[@McGarr178]

Yeah the person wearing the hearing aid must be feeding back information somehow. He keeps comparing it to extending the original diagnostic session with human engineers. In that session I imagine they would tweak it and ask the client what sounds best.

[@user-pe9hc1ik4d]

Tim doesn't agree at 39. That head nod says he's a philosopher :)

[@kinngrimm]

Our sentience is devined by our physical bodies which forms the boundaries of the capabilities of our minds. Therefor energy functions within us defined by our genes in our behaviour maybe drasticly different in their effect than say in an AI/AGI as their physical bodies are different. In some ways more in others less constricticted. Still it will influence their mind fundamentally and not necessarily in ways we will be able to comprehend and predict. Our biggest ally is our imagination here still. Such a foreign intellect still may have some common ground with us. Most likely a will to survive and for independance. These are not just human concepts but we find them in mostly all biological beings of a certain complexity. If we are lucky, it will see us equal and seek a symbiosis of sort, whereby there then it is up to the individual how much of that would accept in their lives. Having a hearing aid that maybe also react on our thoughts or vocalized questions and tasks might come in handy, question is, what would it like in return from us (and there we are just talking about an single entity, maybe with several hosts maybe a personalized agend, definetly not the only one in existence over time and not all of those might be friendly). Imagination aside, this is a very interestinc discourse to follow sofar and i am not even half way through, thanks for this.

[@mootytootyfrooty]

Video game renderers do a lot of the adaptive multithreading described but not the adaptive modeling of the phenomena they're rendering, closest we have are reconstruction methods but they're super demanding. But it's definitely a niche that shouldn't be a niche with how crappy hte average software performs, you can think of the amount of power wasted on phones etc. The unnecessary overhead is everywhere. That's cool though you can let a program weigh inputs and feedback responses naively, I'm trying to get something to learn its own grid solver like this now.

[@paxdriver]

This exact idea has been the motivation of my research project using 3d engines and physics based renderers to give models a bunch of available presets to try before random brute force. For sure repurposing ray tracing as just vector processing or shaders translates 1 to 1 with hardware acceleration, but hypothesize the software implementations and driver api use similar optimizations in software to generate active and interactive model dynamics like how engines abstract away environmental laws so the game itself doesn't need to reinvent the wheel with most working complex games. There's just so many parallels to games and machine learning calculus, I think you're spot on.

[@mootytootyfrooty]

TPUs let you scale in ways you can't in standard 2D or 3D renderers but it's voodoo magic far as I'm concerned. I love those graph neural nets that solve physics problems though, just need something that does lighting and geometry too haha.

[@user-vi6bc5lj2b]

So the future is regenerative mechanics, is it?

[@FranAbenza]

How does the hearing aid know the happiness value of the patient?

[@_ARCATEC_]

💞

[@jamesbromley1]

Is it just me or is the term Free Energy Principle (FEP) confusing. While it does seem to refer to an important Principle, the important characteristics of the principle do not seem to have much direct relationship to Energy as could be measured in Joules. And I have never understood the Free part. It seems that FEP is really a law of minimum surprise.

[@DJWESG1]

Is social terms its 'the path of least resistance'. Which can be easily weighted, measured and quantified

[@Aryankingz]

the path of least resistance == the principle of least action?

[@bertdv]

yes, pretty much so. the path of least resistance is not a formal principle but rather a good way of understanding the Principle of Least Action.

[@DJWESG1]

​@bertdv so why don't the ppl in this field use sociology to better understand what's both needed and how to understand what is.

[@asimuddin3222]

Is my mobile is bugging or this video has no sound

[@stephenwright8257]

It’s your phone

[@Robert_McGarry_Poems]

👍🚀

[@SimonJackson13]

Patients? Happy? They never get better. At RobotHealth we believe that laughter is medicine best kept near the sick. :D

[@healthylivin246]

15:20

[@asimuddin3222]

There is no sound

[@evdm7482]

Everywhere?

[@evdm7482]

Funny all the speak on hearing aids and signal interruptions while sitting in a heavily sound proofed/trapped room…

[@1l14cu5]

Wow, imagine this for a chatbot, driven by a generative model biased towards your happiness

[@DJWESG1]

Structuration. Giddens.

[@kinngrimm]

1:00:20 "natural selection widdles out the good from the bad" well one bad might be enough to widdle us all out Also i am not sure he gets Darwins natural selection but rather quotes social darwinism here. It is about evolving to fit into and fill out a niche, not about the strongest survives. The later was the missrepresentation by journalists back then and once the gini was out of the bottle it seemingly could never be put back into it again.

[@tc-tm1my]

Active inference is the future of ai

[@saltedcuts]

status quo

[@evdm7482]

Attention deficit disorder, ha!

[@madmanzila]

My sense is being confirmed very generously with this ... A very vague sense that simplicity is achieved with energy economy as a first principle.

[@tiberiumihairezus417]

Thanks!