Brain Criticality - Optimizing Neural Computations
Вставка
- Опубліковано 1 сер 2024
- To try everything Brilliant has to offer-free-for a full 30 days, visit brilliant.org/ArtemKirsanov/.
The first 200 of you will get 20% off Brilliant’s annual premium subscription.
My name is Artem, I'm a computational neuroscience student and researcher. In this video we talk about the concept of critical point - how the brain might optimize information processing by hovering near a phase transition.
Patreon: / artemkirsanov
Twitter: / artemkrsv
OUTLINE:
00:00 Introduction
01:11 - Phase transitions in nature
05:05 - The Ising Model
09:33 - Correlation length and long-range communication
13:14 - Scale-free properties and power laws
20:20 - Neuronal avalanches
25:00 - The branching model
31:05 - Optimizing information transmission
34:06 - Brilliant.org
35:41 - Recap and outro
The book: mitpress.mit.edu/978026254403...
REFERENCES (in no particular order):
1. Zimmern, V. Why Brain Criticality Is Clinically Relevant: A Scoping Review. Front. Neural Circuits 14, 54 (2020).
2. Beggs, J. M. The criticality hypothesis: how local cortical networks might optimize information processing. Phil. Trans. R. Soc. A. 366, 329-343 (2008).
3. Beggs, J. M. The cortex and the critical point: understanding the power of emergence. (The MIT Press, 2022).
4. Heffern, E. F. W., Huelskamp, H., Bahar, S. & Inglis, R. F. Phase transitions in biology: from bird flocks to population dynamics. Proc. R. Soc. B. 288, 20211111 (2021).
5. Beggs, J. M. & Plenz, D. Neuronal Avalanches in Neocortical Circuits. J. Neurosci. 23, 11167-11177 (2003).
6. Avramiea, A.-E., Masood, A., Mansvelder, H. D. & Linkenkaer-Hansen, K. Long-Range Amplitude Coupling Is Optimized for Brain Networks That Function at Criticality. J. Neurosci. 42, 2221-2233 (2022).
7. O’Byrne, J. & Jerbi, K. How critical is brain criticality? Trends in Neurosciences 45, 820-837 (2022).
8. Haldeman, C. & Beggs, J. M. Critical Branching Captures Activity in Living Neural Networks and Maximizes the Number of Metastable States. Phys. Rev. Lett. 94, 058101 (2005).
9. Beggs, J. M. Being critical of criticality in the brain. Frontiers in Physiology.
Derivation that only power laws are scale-free: • Fractals and Scaling: ...
CREDITS:
Icons by biorender.com
Brain 3D models were modeled with Blender software using publicly available BrainGlobe atlases (brainglobe.info/atlas-api)
Ising model zooming animations: • The Renormalisation Group
This video was sponsored by Brilliant
To try everything Brilliant has to offer-free-for a full 30 days, visit brilliant.org/ArtemKirsanov/.
The first 200 of you will get 20% off Brilliant’s annual premium subscription.
I am the first here and I am debating on clicking for some reason lol
This only shows 7 days (even with your code) (?)
@@snakejuce Hmm, that's weird. I'll contact Brilliant to double-check this
@@ArtemKirsanov No worries, just thought I'd let you know.
@@ArtemKirsanov ttttttftttttftttftttt
I'm a computer scientist but I really really really love these videos, keep up the good work man
This half-way point between stasis and chaos is also where "life emerges". If you think about life as replicators they need a way to grow and replicate which requires that their lego-blocks should be able to be dis-assembled and assembled. At the right temperatures things are stable enough so that you can keep some information going, but unstable enough so that growth and evolution and "processing"/"thinking"/"natural selection" can happen. I am thinking though that the life emergent point might be based on on covelant bonds on the Earth temperatures but on Mars they might be based on cooler hydrogen bonds as on the Earth covelant bonds are at the critical point allowing photosythesis to create them and digestion, rotting, growing, etc... to repurpose them while on Mars covelant bonds are in stasis so the critical point will be in intermolecular or hydrogen bonds.
I'm also a computer scientist and I like psychology and these kinds of videos.
Samme :))
Dr Leon Chua calls this the edge of chaos. I liken it to a stage microphone on the edge of feedback from hearing its own output from the speaker.
Building networks of these things has got to do some interesting stuff, right?
What was new for me was how the model discovers the geometry of the overall organization, not just pairs leaving identical but increasingly sharp footprints. That’s really nice and rings lots of bells for me.
Absolutly facinating. I am a Machine learning engineer and I could not stop thinking how this knowledge and intuition based on it might be transferred to ML.
Do certain ANN models run near critical points?
I don't think standard ML can implement criticality. I'm looking towards Spiking Neural Networks / Neuromorphic models as the prime candidate for this type of behavior.
ChatGPT4o's response:
The concept of criticality in brain function, as shown in the UA-cam video screenshots, can be applied to machine learning (ML) algorithms in several ways. Here are a few ideas:
1. **Dynamic Parameter Tuning**:
- Use principles from criticality to dynamically adjust hyperparameters in ML models. For instance, a system can be designed to detect when the model is near a critical point and adjust learning rates, dropout rates, or other hyperparameters to optimize performance.
2. **Spiking Neural Networks (SNNs)**:
- Implement Spiking Neural Networks, which are inspired by how neurons in the brain communicate. These networks can operate near criticality, offering potential improvements in efficiency and robustness.
3. **Self-Organized Criticality (SOC)**:
- Integrate self-organized criticality into ML models. This concept can help in maintaining a balance between stability and adaptability in neural networks, enabling better generalization and avoiding overfitting.
4. **Criticality-Based Regularization**:
- Develop regularization techniques based on criticality to prevent overfitting. By encouraging the network to operate near critical points, it can achieve a more balanced learning process, improving both training stability and generalization.
5. **Adaptive Architectures**:
- Create adaptive network architectures that can reconfigure themselves based on the critical states detected during training. This could involve changing the number of neurons, layers, or connections in real-time to optimize learning and inference.
6. **Energy-Efficient Computing**:
- Leverage criticality to design energy-efficient ML models. By mimicking the brain's energy-efficient processing near critical points, ML models can reduce computational costs and power consumption.
These methods aim to make ML systems more efficient, adaptable, and closer to the natural intelligence processes observed in the human brain.
This ties into the weight initialization of layers in deep neural networks in machine learning. If the magnitudes of the weights are too small then the outputs diminish with each layer, otherwise if the magnitudes are too great then the outputs blow up. Balancing these weights allows for the stacking of many layers which has enabled the great progress we have seen in deep learning in recent years.
I thought exactly about the same thing. This is the vanishing or exploding issue in the forward/backward pass in ANNs. To alleviate this problem, there is also batch normalization which helps keeping the activations std to 1 throughout the training process. The skip connections also help keeping the flow of information. I also thought about the attention mechanism used in transformers. For each output, it takes the weighted average of the input tokens. These positive weights add up to 1 thanks to the use of the softmax function, keeping the flow of information constant through the layers. Transformers combine all these tricks (they use layer normalization instead of batch normalization, but the idea is the same).
Moreover, the original problem solved by the attention mechanism used in transformers was that the hidden state in RNN/LSTM acting as a memory state hardly retained all the information of the sequence of tokens that was previously processed. The information about the past tokens sort of vanishes (or at least is incomplete) as the model goes forward through the tokens. The attention mechanism serves as a kind of skip connection that allows the model to look at all the previous information which is then preserved and can flow much more easily. In the end, even in ANNs, good information flow is central to their proper functioning.
Now, it would be very interesting to know how nature came up with a good information flow management in the brain. The critical brain hypothesis is interesting, but it seems to me that it only makes some observations related to the critical phenomena but doesn't really explain the mechanism causing this criticality (it might very be the ultimate goal of neuroscience). Researchers in AI could then take inspiration from it.
@@chocochip8402Indeed. Discovering the cause of criticality will be the end of neuroscience, and the death of God as well.
This is one of the best videos I've ever come across in something like 10 years using this platform.
I can't overstate how good this was. Amazing job, I'm looking forward for your future content
Wow, thank you so much!
At a long-time and large-size scale water is at a critical point on the earth (in that it is in liquid, gas, and solid state). However, more importantly carbon-nitrogen-oxygen covelant bonds in life are at the critical point in long and short time scales, allowing its bonds to be repurposed and allowing self-replication and evolution. On Venus these bonds are unstable, while on Mars these bonds are at stasis. I think on around Mars/Europa hydrogen bonds may be at the critical point so you might see complex "ice crystal" life while on Venus some sort of weird sulfuric acid compounds are at the critical point.
Every video you have made so far is a masterpiece. You cover a wide variety of computational neuroscience topics from place cells to wavelets; with each topic covered in exceptional detail. You are able to convey abstract topics in an intuitive and visual way that is unparalleled.
Keep up the great work man
No one explains better than you do. I knew all these stuff in their separate domains, but I've never truly understood the connection as I have now. When at 25:07 you justified the passage between electrodes and neurons it blew my mind of pure happiness!!
This is the most amazing video I have seen on UA-cam for a while. This is Science Communication at its best. Thank you so much!
Completely agreed 👌
Wow, thank you so much!
Wow. Self-Organized Criticality. Scale invariance of Relevance Realization. Deep-continuity hypothesis. Our metabolism powers our virtual engines which are optimized and orchestrated on top of the background "hum" of critical neural objective reduction. Thanks for this great work.
Kiváló előadás a lényegről.
Nagyon jó oktatási anyag, kutatóknak is javasolható.
Köszönet érte!
Dear Artem, thank you for this glorious video! Well made and inspiring! You triggered another neural avalanche of excitement in me! My brain transitioned from rapid eye movements and sleepiness to the rabbit hole of self-organized criticality!
Amazing video! I did an undergrad research about brain criticality. The idea was to create an analog of the connectivity matrix for the Ising model in the critical temperature to check if the graph topological properties match with the ones measured in the resting state with fNIRS.
Oh that's sound great! There's something you've published?
@@danyielsanchez5159 Not yet
This is beautiful. I am interested in seeing the effect of psychedelics on control parameters.
Thank you! Interesting thought indeed!
I had a powerful realization during a deep trip where I realized that life and consciousness are the result of the feedback/recursive character of the critical line. The more you can tune toward greater coherence, the higher the degree of consciousness.
your intuition is right, read: Carhart-Harris, R.L., 2018. The entropic brain-revisited. Neuropharmacology, 142, pp.167-178.
guy who's fried his brain with psychedelics: WOAHHH BUT WHAT IF HE WAS ON ACID MAN
@@philipm3173 holy fuck I did that on weed but I failed to realize the second part.
Are you kidding man? On a road trip rn and have been talking about this with friends. Can’t believe this just came out, very excited to listen!
This is so so so well made! It makes you feel as if you're gradually discovering these results for yourself and it feels fantastic doing so!
Your videos are truly a gift! Amazing research and video quality. Keep it up!
OUTSTANDING video! :D
You taught the concepts in a very clear way and the animations are simply insane. I love it!
Another fascinating video, Artem. The work you’ve put in to making the material accessible to non-specialists has definitely produced a pedagogical jewel. Amazing.
Artem you are a genius! Your videos made me interested in neuroscience and now I am fully devoted to reading about it. I recently read about criticality and now I see your video and it's just so beautiful. I wish you talked about self organized criticality too
This is the best UA-cam video I have ever seen. You explained everything masterfully! Thank you for giving my curiosity a vision, I’m so excited to explore more.
Studying the Ising model for my thesis right now. I never would have thought that there is a connection between the model and NN's (which also feels extremely natural). Nice content
This might be my favorite Artem Kirsanov video. A masterpiece of masterpieces. Thank you so much for making these.
Very impressive visual animations. Helped a lot with understanding the concepts
This is SO well done. Scale-free avalanches in the brain makes perfect sense, since we are trying to self-resonate, such that information is not lost as it echoes up and down the various physical thresholds which constitute our brains from atoms all the way up to the whole structure.
Despite these epiphanies handed to me on a silver platter, I'm still having trouble wrapping my brain around how any of this helps keep neural networks in a state of unstable equilibrium, what are the hidden variables that prevent self feedback oscillations from getting phase locked much like a seizure, or descending into complete chaos? It's much reminds me of a table full of pendulums that stand upright when the table is randomly vibrated but much more complicated.(because they're all connected to the same table they want to sync up, because the vibration is random they seldom do, yet within the narrow range of vibration they all stand up!)
@@petevenuti7355 You have to remember our brains, like the rest of us, evolved naturally. Therefore the near-critical point is a universal feature of life. Imagine you want to farm entropy, where do you go? You go where it’s being formed, at the edge of a phase transition - kinda like how we build along coastlines, or better yet how primordial life still clings to hydrothermal vents deep underwater. The transition from eddies to flows is where all the magic happens.
In the brain then there are feedback systems preventing your bad feedbacks, because it’s actually designed around physical minima, carved a home in energy gradient which is stable despite its complexity - life is a self-stabilizing dissipative structure, using the pull of entropy to orbit equilibrium.
@@anywallsocket "life is a self-stabilizing dissipative structure, using the pull of entropy to orbit equilibrium" what an interesting way to think about it.
@@anywallsocket That's beautiful, thank you
@@anywallsocketbeautiful
Awesome! I'm going to recommend this channel to my Neuroscience class.
Perfect, I've been reading connectome harmonics papers recently so this is very much topical to me.
You have a talent of combining beauty and science. These are often thought to be separate; thanks for illuminating the bridge.
Your videos are always enlightening; thanks for the consistently great content!
finally someone talking about phase transitions
This is a great topic and a beautiful presentation based on a great paper. Excellence all around.
The insight, that cyclic relations define the geometry of the map, is a nice key insight breaking out of simple Pavlovian association lists.
Always a joy when Artem drops a video!
This Video is so so wonderful, thank you!! All very beautiful, interesting and clear. Good luck for next videos and thank you
Such an intricate and complex topic, so well explained. Truly remarkable!
Thanks for leaving sponsor at the end. I watched the whole thing.
Brilliantly explained. Please carry on making this type of videos.
This video is so interesting. Thanks a lot for making this video and please keep delivering content about computational neuroscience in an informative yet easily digestible way!!
My first time viewing. What an excellent job. Simply correct in matters, meaning and math. I am very impressed.
I was sick today and binged some of your videos. So far, they're all brilliant and I love the aesthetic and craftsmanship you put into them. I thought of the Ising model as you were talking about phase transitions, and then you bring it up -- truly comprehensive and love that you are bringing physics into your videos! Super interested in similar systems, like Kuramoto oscillators which can possibly describe large scale brain oscillations, and which have mathematical similarities to Bose-Einstein condensates.
This is one of the most thought provoking videos I have ever seen. This is now one of my favorite channels.
I am experimenting with spiking neural networks evolved through indirect encoding and i experienced spike wanishing in the past. This video blew my mind and i've learned a ton from it. I'm super inspired. Thank you!
Fascinating and incredibly well put together video!
Bro this video is just outright phenomenal . Thank you for your time
This is fascinating work and you explain it perfectly! Thank you!
This video helped me get dangerously close to thinking I understand the nature of the universe and myself inside it. Thank you for making such a brilliant video that's available for everyone to learn from.
Exceptional video. Thanks for putting in what I'm sure was a monumental amount of work to explain several quite complex concepts clearly and concisely. Subbed!
Fascinating. Thanks for making this.
Exceptional work explaining and visualizing this fascinating topic! Thank you from the bottom of my heart for gifting us your videos ♥
Thank you for posting this. I've been trying to find new ways of explaining the 'grokking' behavior of ML, and how this is a phase transition behavior similar to Flory-Huggins, liquid crystals, weather patterns, etc. but have not had a good way of describing it beside vaguely grasping at Fourier decomposition of a signal. This is a more detailed overall explanation. Glad it also applies (as expected) to biological neurons. Best wishes.
I think your video inspired me to how to solve a problem in my research project, about the optimization in critical stage, and the communication by long-range coupling. Thank you!
U dont know shiiiiit u are talking about 🤣.....samo rokni malo magnezijuma i malo cinka...odma ti bude bolje 🙃
This work of art is as valuable as works of Plato. Thank you for bringing to our consciousness
Please do an analysis of the renormalization group. Your exposition of critical phenomenon and self-similarity is extremely elegant and intuitive, beautiful work!
Wow! This was the best video I've seen for a while!
And it gave me an idea about how this ideas described here that can have a huge impact on Graph Neural Networks!
Thanks for such an amazing content!
you explain concepts so well & eloquently. the theoretical simulations, etc.
This video is extremely well done! Thank you!
Fascinating. I come from math, and in very abstract algebra and geometry there are several notions of dimension that emerge from observing some power law. The object of finite dimensions are, of course, of the particular interest.
I got interested in neurology a few years ago But lost interest. But this video Has definitely made me want to study it again. You explained everything so simply and perfectly. Definitely one of the best Scientific videos I've ever seen on UA-cam❤❤❤❤
THANKYOU so much for scale invariance.
One of the best videos I’ve seen on UA-cam! (The others are also your videos)
Amazing video!
I saw this talk related to neurofeedback and your video helps to understand it better. I plan on picking up a copy of the book. Thank you.
Tuning Pathological Oscillations with
EEG Neurofeedback and Self-Organized
Criticality - Tomas Ros
Hey Artem
Very nice video, i have been doing Percolation models for physical systems for a while. It is rare to get percolation lattice simulations on youtube outside of very esoteric channels that nobody knows of.
It is interesting how it can be mapped to Neuroscience.
10/10
This is super-high quality content ! Congratulations !
Damn, that was eye opening! Thank you for making this!
Thank you for this video, I just learned about some of these concepts without knowing any of this background. Thank you again!
Thanks a lot, Artem....This viedo was awesome
This is one of the greatest channels on UA-cam.
OMG the graphics of this video are just popping off! I absolutely adore the font choice and visualizations. I can't believe you haven't passed 100K subs yet! But I'm sure you'll get their soon, and I'll add a small +1 to that count :)
Awesome presentation !!!
Thank you. Very informative
Amazing content, thank you so much
I am a rather regular software developer but I kind of try to avoid too much math but this video is phenomenal that even with my forgotten knowledge I could easily follow what was explained here.
incredible video, hope you make more like this!
What a beautiful video !
What a great video! Keep it up
One of the most intellectually rewarding videos I've ever seen!
this is literally so good. nice job! i learned so much :)
Thanks man, you helped me to finally understand this stuff.
This seems theory to be resonating with a lot of other fields of science, as well as experience being embodied, and I want to thank you for presenting this topic in such an accessible way! I think that it is important that we continually update our internal models of the world and our self to be able to stay in touch with it.
Artem, man, really great content. Making me want to go into research/industry neurosci or neuromorphic computing.
Great content. Thanks.
Thank you. Very, very interesting.
Artem, you do sci comm like no other. Thank you 🙏
Well done!
Just wow. Amazing presentation, great content.. I'd love to work with you some day
Thank you.
22:48 you absolutely just blew my F-ing mind.
This is reminding me of the book.. The computaional Beauty of Nature. Great work.
This is so well explained and an amazing video!
Amazing work, thanks a lot!!
Geoffrey Hinton has developed a forward forward algorithm for learning. Essentially there is an awake and sleeping phase both required for learning.
This videos are amazing! Thanks for giving this content for free. I would be really interested on a video about the Free Energy Principle by Friston.
So cool, thank you.❤
Amazing video! As an undergrad, I greatly appreciate your videos!
You are awesome your lectures are excellent work
I'm studying chemical physics. The first half is soooooooo clear! Thank you
This neuroscience video is probably the best explanation on the Ising model I’ve seen!
Thanks! :D
This is true. Although I missed a word that the Ising model stands out in that it can be solved analytically.
I'm a computer so I really really really love these videos, keep up the good work man
Beautiful visualizations.
Such quality content 👍👍👍
I know you know but your videos make real intellectual satisfaction because they are sooooo great
Man. Just... thank you!