How realistic are modern neuron models?

[Aidyan]

Nowadays we hear a lot about AI deep learning artificial neural networks. People speak about simulations of billions of neurons and even much more synapses. However, as far as I understand it, the artificial neurons used in these simulations are extremely simplified models of real biological neurons. The question is how much simpler? How far away are we from being able to simulate every neuron realistically? Can we simulate today exactly at least one single neuron exactly? Do we know for sure everything about how neurons work?

 

[rootone]

The internal structure and chemistry of an individual neuron is quite well studied and understood.
How these form networks which can trigger whole body responses to external stimuli, or represent memories, or give rise to conciousnes is not well understood.

 

[jim mcnamara]

It is a computer emulation - it tries to duplicate a few aspects of how neurons interconnect, not how individual neurons do their thing biochemically.
From https://www.doc.ic.ac.uk/~nd/surprise_96/journal/vol4/cs11/report.html

- note the word inspired, in the quote below, not a duplication in any sense.

An Artificial Neural Network (ANN) is an information processing paradigm that is inspired by the way biological nervous systems, such as the brain, process information. The key element of this paradigm is the novel structure of the information processing system.

So your answer is 'not close' We have some PF'ers who are in the field, both in neural networks and in neuron biochemistry. If you meant can we make a single neuron (functionally) cloned in a computer, then I do not know.

Answer to last question: No, we do not know everything there is to know about neurons

 

[Aidyan]

So, as far I can understand that, present neural network simulations never will work in producing, say self-driving cars or some other sort of intelligent machine, since they are based on building blocks which are much too primitive and get not even close to real neurons. Is then all this discussion on AI just a media hype?

 

[jim mcnamara]

If AI problems were super simple, we would have had totally safe self-driving autos for about 15 years. That is when the first robot car challenges were run.
Does that answer your question?

I am in no way demeaning AI efforts, it is just a set of hard problems. Lots of progress - depending on how you want to view it. Some of the current progress is hardware related as well.

OR you could take Randal Munroe's point of view:

https://imgs.xkcd.com/comics/self_driving.png

 

[Laroxe]

This might be an interesting starting point, its written by Mark Humphries who holds the Chair in Computational Neuroscience at Nottingham University and it reviews some of the current thinking about neurones in the brain. There is still a lot of work coming out suggesting even more complexity in functioning.

 

[Aidyan]

This what? I see only a white patch. Link missing?

 

[Laroxe]

This what? I see only a white patch. Link missing?

Couldn't get the link to work, wouldn't even let me delete this one. I tried all sorts and I am hesitant to try again. the article was on medium.com/the-spike - and was called - /your-cortex-contains-17-billion-computers-9034e42d34f2
putting the two together should work. Sorry.

 

[Aidyan]

I found it here: link
The bottom line:

"It suggests that deep learning and its AI brethren have but glimpsed the computational power of an actual brain." - "But if we think the brain is a computer, because it is like a neural network, then now we must admit that individual neurons are computers too. All 17 billion of them in your cortex; perhaps all 86 billion in your brain. And so it means your cortex is not a neural network. Your cortex is a neural network of neural networks."

We are light years away from what the overhyped AI sensational-sounding article titles would like us to believe.

BTW, the above self-driving.png of three years ago was prophetic: https://electrek.co/2020/06/25/tesla-autopilot-confuses-burger-king-stop-signs-ad-campaign/

 

[madness]

Nowadays we hear a lot about AI deep learning artificial neural networks. People speak about simulations of billions of neurons and even much more synapses. However, as far as I understand it, the artificial neurons used in these simulations are extremely simplified models of real biological neurons. The question is how much simpler? How far away are we from being able to simulate every neuron realistically? Can we simulate today exactly at least one single neuron exactly? Do we know for sure everything about how neurons work?

There are essentially different fields trying to understand neural networks (both biological and artificial) and single neurons. Most would agree that it is not worth trying to put very detailed simulations of single neurons into a network model. Even if you could do it, you woudn't achieve anything useful. To quote Norber Wiener:

"The best model of a cat is another, or preferably the same, cat."
https://www.jstor.org/stable/184253?seq=1

In other words, if you build a realistic model of something you don't understand, you will just have a model that you don't understand of the thing that you don't understand.

But the short answer is no, we don't understand how neurons work (we don't even understand how ion channels work). We also don't understand how networks work, but that's not because we don't have complicated enough neuron models - we don't understand them when we use simple neuron models either.

 

[madness]

So, as far I can understand that, present neural network simulations never will work in producing, say self-driving cars or some other sort of intelligent machine, since they are based on building blocks which are much too primitive and get not even close to real neurons. Is then all this discussion on AI just a media hype?

AI systems based on the simplest artificial neurons vastly outperform those based on biologically realistic neurons. This isn't the problem that AI is facing.

 

[Jarvis323]

I think that the main debate as to whether classical artificial neural networks could match the power of a real brain, is as to whether consciousness and the abilities of a real brain necessarily rely on quantum mechanical processes. Most notably, Roger Penrose, the guy who just won the Nobel prize in physics, is a proponent of this theory. He and Stuart Hameroff propose that "consciousness might be the result of quantum gravity effects in microtubules" within brain cells.

https://reader.elsevier.com/reader/...18C7B3213723F1D742F97C3786C5A08215D05715EE4FA

https://www.discovermagazine.com/th...n-consciousness-one-scientist-thinks-it-might

Penrose also asserts that consciousness cannot be explained without the brain exploiting quantum effects. As brilliant as Penrose is, most computer scientists and physicists think that his assertion is unfounded. It is mainly his reputation that really carries the theory.

Wikipedia summarizes his main argument as follows:

He bases this on claims that consciousness transcends formal logic because things such as the insolubility of the halting problem and Gödel's incompleteness theorem prevent an algorithmically based system of logic from reproducing such traits of human intelligence as mathematical insight.

https://en.wikipedia.org/wiki/Roger_Penrose#Physics_and_consciousness

I personally wouldn't discount the possibility that the brain non-trivially exploits quantum effects. I have an interest in quantumn biology in general, but some of his arguments seem a little naive to me.

Scott Aarronson (a leading theoretical quantum computer scientist) has weighed in on the topic extensively.

https://www.scottaaronson.com/democritus/lec10.5.html

Of course, even if Penrose is right, who is to say we can't eventually replicate the "microtubules that exploit quantum gravity" in an artificial network.

 

[atyy]

We are light years away from what the overhyped AI sensational-sounding article titles would like us to believe.

We may be, but those who propose the reason is insufficient complexity in each neuron have not shown us how that would solve the problem. So it is not clear that that is the correct reason for skepticism.