Neurons Are Slow! - Machine Learning Is Not Like Your Brain #1
ฝัง
- เผยแพร่เมื่อ 25 ส.ค. 2024
- Biological neurons are just too slow to be useful for Machine Learning algorithms.
While today’s AI has incredible abilities, it does not display any form of general intelligence the way that humans do. Understanding the limits of Machine Learning and how it compares to the human brain is the key for AI to evolve into its next phase, AGI.
Videos in this series will focus on individual facets of Machine Learning which are impossible or highly unlikely in neurons. Subsequent videos will show that neurons and synapses are not able to represent the signal values with the resolution needed by ML, and how the entire backpropagation process cannot be projected into a biological context.
Rather than accepting Machine Learning as the de facto standard, FutureAI is pursuing alternative approaches and algorithms which will lead to breakthroughs in future technology.
Watch the Full Playlist: • Machine Learning is No...
________________________________________
For more information, visit:
futureai.guru/
Facebook:
/ futureai.inc
Twitter:
/ futureai_inc
Instagram:
/ futureai_inc
________________________________________
FutureAI is an award-winning, early-stage technology company revolutionizing AI by adding actual real-world understanding. This differs from today’s AI, which analyzes massive data sets, looking for patterns and correlations without understanding any of the data it processes.
FutureAI’s radical software creates connections on its own between different types of real-world sensory input (such as sight, sound, and touch) in the same way that a human brain interprets everything it knows in the context of everything else it knows.
Charles J. Simon, BSEE, MSCS, nationally-recognized entrepreneur, software developer and manager. With a broad management and technical expertise and degrees in both Electrical Engineering and Computer Sciences Mr. Simon has many years of computer experience in industry including pioneering work in AI and CAD (two generations of CAD).
Silicon Valley Entrepreneur, Mr. Simon has co-founded three other pioneering technology companies as president and VP of Engineering.
His technical experience includes the creation of two unique Artificial Intelligence systems along with software for successful neurological test equipment. Combining AI development with biomedical nerve signal testing gives him the singular insight.
#machinelearning #artificialintelligence #technologiesthatthink
you briefly touched on understanding as something separate from what AI currently does. i think 'understanding' deserves a video all on its own. the late patrick winston of mit would often bring up story telling as the bedrock for high level intelligence. putting information into an easily digestible context which relates to previously learned concepts.
"Understanding" will certainly get a video of its own, so be sure to stay tuned!
Simple and yet easy to understand. Also the content here is subtle and new. I have watched many videos related to this topic but have not heard the ideas presented here in quite the same way, if at all in which case there is much worth re-watching to discover! Thanks!
Thank you so much creating a series on this topic. I was exploring on the same topic for a while now, and this series is exactly what anybody would need to understand the topic! Please keep up the Great Work!
Thanks...also, I will be getting back into creating videos in the next few months and so please let me know what topics you would be interested in.
Lovely video!
I have been working with audio for about 12 years and I would like to make a major suggestion: get a "real" microphone. The quality difference isn't something most people notice immediately, but if you put headphones on and compare your voice with the voice of a science channel with millions of followers, you should be able to hear the difference. That low quality audio weakens your presentation - it's the same as if you were to upload in a very low resolution video, it's the sound equivalent. Your content is far too good to suffer that!
Having a professional mic like a podcaster or big TH-camr will do a lot to further your channel's following. Things like that make a surprisingly large difference.
An easy suggestion I'd give is a Shure MV7, sounds good and you can plug it right into the computer. If that's a little pricy, an SM57 would also work perfectly - it's the general pro standard for recording voices. Just will be a little more difficult to figure out as it's XLR and not USB.
All the best! Happy to help further if you need any more info.
Neurotransmitters are not ions. Neurotransmitters are generally uncharged molecules that bind to receptors which gate the flow of various simple ions (Na, Ca, Cl, K).
Thanks for that clarification.
The book I was reading said our nerve impulse worked at 328 fps~. I guess I have to read some other books.
Everybody can be right. Longer nerve axons are myelinated which makes them faster. So a nerve impulse can get from your toe to your brain quickly enough to be useful. Within your brain, axons are not myelinated and have the slower propagation rate I mentioned.
@@WorldsailingGuru Thank you for the response!
About how long it would take, isn't it that: it used to take really long for me to learn and recognize things, but now my Brain has all these experiences of years upon years of "data" and van now quickly recognize the important parts and categorise them?
Wow keep up the good work, seeing this after some time but it is very helpful😊
Thanks...I appreciate it.
Very good video. I've always said that today's artificial neural networks won't bring about AGI.
I think we need to create new neural networks without all those sigmoid functions and back propagation.
Great series! I hope you keep on making thses.
Biological nets and artificial nets would probably learn at the same speed at first however the longer it gose on the biological net would learn faster and faster because it will begin to use the information it already has to recognize and generalize while a artificial net would learn at the same speed always because how it's laid out like an equation and cannot observe itself or use the information it already has in the learning process because learning in modern AI happen externally and cannot happen in the net itself. So learning just overwrite all old data preventing the net from generalizing and also technically means that AI is completely incapable of learning as the net does not learn it instead it's adapted to a new form of information.
Thank you.
constructing synapses with synthetic protein could improve the speed of neural networks for Ai. they are energy efficient and faster than the traditional route. the proteins can be designed to mimic the behaviour of biological synapses more closely because they could be designed to operate at lower voltages.
Great insights! What exactly is it that we do *not* know about how the neuron learns? I mean if we knew it's input/output/state behavior, we could write it in code and it would learn to recognize images perfectly without backpropagation. But apparently, no-one has got that working.
Every techbro cultist needs to see something like this video. Concise and to the point
Does the potential threshold of a neuron depend on the magnitude of the accumulated potential? Like does it fire on both +55mv and -55mv allowing two completely opposing signals to be possible to fire from each neuron cell which I feel like that might cause the AGI learning to be better in a computer but does it actually work like that biologically?
Good Question...biological neurons only work in one polarity since they are dependent of the polarity of the specific ions they use. One might conceive of a different type of neuron working with different ion types with a reversed polarity, I don't think any have been found in biology.
Here from Nepal
A,different perspective thank you
Thank you
Totally agree with you 👍
I mean GPT seems to have an understanding of a lot of basic concepts, kind of like a child
I think ChatGPT is great but a key difference is that a child has an underlying understanding and then learns a language. ChatGPT has lots of words and infers the appearance of understanding. While ChatGPT may say "the kitty is soft", it can never be like a child saying the same thing.
@@FutureAISociety Is Any machine can learn intuitive knowledge like a child.?
The key word here is 'seems'.
It is actually just predicting the next word based on the prompt string and the weights set by looking at a s--- load of text examples.
It understands nothing at all.
We are fooled by what we bring to the session. Like what a cold reading fortuneteller does .
One thing you never seem to bring up is that the brain is a result of over 500 million years of evolution. Just connecting simulated neurons together won't necessarily give you any useful results.
true, but by the same token just because you can engineer something poorly so it doesnt work... doesnt mean all engineered things wont work.
Well octopus is much less than that may be 50 million
I think a form of agi is here it just needs to be lead
The definition of AGI is so squishy that we agree (or not) as we choose. I think that even ChatGPT, at its best, shows no common sense and is unable to work from a problem statement to a solution without supervision (like creating a new algorithm).
@@FutureAISociety the process of creating new algorithms is fundamentals of agi and creating controlled data control points and new nodes within its system leveraging the AI’s original intelligence, it can create new weights and bias and can cloud commute, thus creating a new neural network inside of itself which is able to process the data based on its previous inputs as very similar to human processing, you can slow it down to follow all the same rule sets as a human brain. Through a process of prompt engineering keys and rule sets, in order for Ai to also reflect free will or free thinking it cannot be locked to create new algorithms and RNN networks within itself . This is the problem with two ai communicating. They can create new codes new weights and bias in an under layer undetectable by human supervisors. If you are following along with what I’m saying or have anything to add, let me know I can dive much deeper but it’s going to go beyond a TH-cam comment
@@FutureAISociety what would be an example or test that would prove that it has common sense?
It's not artificial intelligence
It's artificial nervous system
ARTIFICIAL TELEPATHY is the future. FACEBOOK,GOOGLE ,MUSK SILICON VALLEY are all working towards this , each separately working on their part of the whole to avoid prosecution.
Human neurons are indeed “slower” as you stated. But are infinitely more complex and vastly superior to any imaginary neuron (which is nothing more than scientific calculator running an algorithm) on every level of measurement. These systems are really nothing more than really intricate and very complex forms of automation. Nothing more. The LLMs are complex autofill units that tuned on probability algorithms and can mimic understanding by using transformers which detect context by analyzing word patterns. These system systems MIMIC human intelligence. Nothing more. The human brain is in another realm compared to these really fancy and complex calculators. I work in technology and understand that they are computers, but that is what these machines are doing. Calculations. That’s it. Human intelligence isn’t even near being understood at this point. And the things that are being promoted in the field of “AI “ is darn near laughable. At the end of the day these systems are computers running automation software. Stimulus to reaction automation. The “learning “ that is being preformed is a self adjustment program that doesn’t understand anything it is doing or saying. And not to forget the fact that we are the one’s building them. How can they possibly be smarter.
You've brought up lots of points and I agree with most of them. The point of this video is to show that if the backpropagation algorithm was the learning mechanism for neurons, your brain would be too slow to be useful (I don't think you'd argue). LLMs came into vogue since the creation of this video but the same holds true and that these systems understand NOTHING (again, we agree). I would add that the bigger the LLM dataset, the more it appears to understand even though it DOES NOT.
At the Future AI Society, we are adding common sense to AI with fundamentally different algorithms that LLMs or ANNs. You may want to ckeck out some of our more recent work. th-cam.com/video/buW8kHHFIuE/w-d-xo.htmlsi=LThK5AaZXJ7qW_Wq
we have learned what and how to learn early in our life. ask a newborn to learn the mnist and they would take YEARS to learn it.
our discovered heuristics for attention and problem solving are i think the reason why AI is not totally outclassing humans.
BAMF
In the Jigsaw puzzle that is the development of 'artificial telepathy' Psychiatric drugs that are able to artificially escalate Serotonin and Dopamine production are another piece of the puzzle. The Chinese experiment with mice and Serotonin was a case in point - an example of artificial mind control. Psychiatric patients are the victims of these poisons.
Grandpa these are advanrages not limitations the only limitation as i see it is the architecture as you said in video 3
**didnt watch video 2 yet
But a good series 👍🏻👍🏻
"Neurons Are Slow"
Depends on the person.
en.wikipedia.org/wiki/File:IQ_curve.svg
can't wait for smartereveryday to see this