The problem of designing an AI to understand what it sees

@sonhouse said
@KellyJay
You are just showing how far behind technology you are, Intelligent AI will be made of complex neural nets not 1's and 0's.

Umm., since neural networks are implemented on conventional computers it is all 1's and 0's.

@deepthought said
since neural networks are implemented on conventional computers it is all 1's and 0's.

I think its interesting to note that, in theory, an artificial neural network need not be designed to send 1's and 0's signals because it could in theory be made to work on analogue signals. However, I am currently unaware of any truly pure analogue artificial neural network and it should be pointed out that even in NATURAL neural networks, such as in the human brain, send there signals around digitally and NOT in analogue.
In other words, even NATURAL neural networks, such as in the human brain, send there signals around in 1's and 0's.

@KellyJay

Read the last sentence of my last post.

@humy said
The "answer" to exactly what question relevant to potential AI intelligence?
Or have you just suddenly changed the topic of this conversation?
And what has "moving rocks around the ground" got to do with it?

About the same thing as sending patterns through a processor.

@humy said
I think its interesting to note that, in theory, an artificial neural network need not be designed to send 1's and 0's signals because it could in theory be made to work on analogue signals. However, I am currently unaware of any truly pure analogue artificial neural network and it should be pointed out that even in NATURAL neural networks, such as in the human brain, send there ...[text shortened]... even NATURAL neural networks, such as in the human brain, send there signals around in 1's and 0's.

I'd be willing bet its a little more than that, considering how we process information isn't no where near the same as how it is entered into a computer no matter how advance it is.

@kellyjay said
I'd be willing bet its a little more than that,

no, all the signals in our brain can be viewed as being transmitted in 1s and 0s because either a connection is currently transmitting a signal or it isn't and there are no half-signals. These are just the scientific facts. Computers already do this so this isn't a barrier to computer.

@humy said
no, all the signals in our brain can be viewed as being transmitted in 1s and 0s because either a connection is currently transmitting a signal or it isn't and there are no half-signals. These are just the scientific facts. Computers already do this so this isn't a barrier to computer.

You view them anyway you want, how things are processed isn’t digitally; however, looking at the processes as digital does gives us insight into it.

@humy said
no, all the signals in our brain can be viewed as being transmitted in 1s and 0s because either a connection is currently transmitting a signal or it isn't and there are no half-signals. These are just the scientific facts. Computers already do this so this isn't a barrier to computer.

I don't think this is right I'm afraid. I really don't know much at all about this so I am at huge risk of being wrong here. However, when an axon fires there are numerous ways neurons transmit the signal to each other including a mode of transmission that involves induction between neighbouring nerve fibres, called Ephaptic coupling [1] (it is somewhat controversial). Basically no purely digital model of biological neurons works, see [2] for a list of models for this. The interesting question is whether this matters or not as to how to make artificial intelligences work.

[1] https://en.wikipedia.org/wiki/Ephaptic_coupling
[2] https://en.wikipedia.org/wiki/Models_of_neural_computation
[3] https://en.wikipedia.org/wiki/Biological_neuron_model

Especially note the last section in [2] - Embodiment in Electronic Hardware.

@DeepThought
Sure, they simulate neural nets with 1's and zero's but they are pretty inefficient in terms of energy compared to real brains;

@deepthought said
I don't think this is right I'm afraid. I really don't know much at all about this so I am at huge risk of being wrong here. However, when an axon fires there are numerous ways neurons transmit the signal to each other including a mode of transmission that involves induction between neighbouring nerve fibres, called Ephaptic coupling [1] (it is somewhat controversial). ...[text shortened]... logical_neuron_model

Especially note the last section in [2] - Embodiment in Electronic Hardware.

I had not previously heard of ephaptic coupling because I am an AI expert but not an expert on natural networks and only have pretty basic expertise even on artificial networks as I concentrate all my research and studies on only software for AI and not AI hardware (mainly because I have no real money or means to make hardware but making software is easy because all one needs is to now how to program, which I do). So I take it then that, at least some of the time, some of the signals between natural neurons can be said to be analogue?
By the way; I was only referring to the signals between the neurons and not what is called the "weights" and the "threshold values" within each neuron because I already knew those latter two things can and often are represented with true analogue as opposed to digitally.

@humy said
I had not previously heard of ephaptic coupling because I am an AI expert but not an expert on natural networks and only have pretty basic expertise even on artificial networks as I concentrate all my research and studies on only software for AI and not AI hardware (mainly because I have no real money or means to make hardware but making software is easy because all one n ...[text shortened]... ew those latter two things can and often are represented with true analogue as opposed to digitally.

So I take it then that, at least some of the time, some of the signals between natural neurons can be said to be analogue?

Well, there's clearly going to be some variation in the amount of neurotransmitter in play and the physical conditions of gap synapses (direct electrical signaling) are going to vary. Some in vitro experiments on cat neurons I heard of in the '90s indicated that the signaling is digital, either the neuron fires or it doesn't, with variation in the threshold potential required [1]. But this is an in vitro experiment, when surrounded by the rest of the cat the neuron might behave differently.

If you think of the evolutionary basis for this, I imagine something along the lines of a sensor on some early blob of cells from 500 odd million years ago causing the animal to change shape to move towards the stimulus. This turned into a neuron activating smooth muscle. One might expect the strength of the response to depend on the strength of the stimulus. So really it would be a surprise if neuron responses were purely digital.

Whether this is facilitated by signal type and strength between neurons or by biological analogues to the actuation thresholds and weights that artificial neural networks use requires the opinion of a neurologist.

[1] Sorry, I can't give a reference for this, a biologist told me this over 20 years ago and I haven't a hope of finding a reference.

@DeepThought
Here is one piece by Paul King, a neuroscientist about that digital V analog question:

https://www.forbes.com/sites/quora/2016/09/27/is-the-human-brain-analog-or-digital/#1d1b3fd37106

My expertise lies in digital hardware, and the strict 1's and 0's inherent in such gates, and, nand, nor, Xor and so forth.

But there is another kind called Approximate signal processing:

http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.50.6776&rep=rep1&type=pdf

@deepthought said
I don't think this is right I'm afraid. I really don't know much at all about this so I am at huge risk of being wrong here. However, when an axon fires there are numerous ways neurons transmit the signal to each other including a mode of transmission that involves induction between neighbouring nerve fibres, called Ephaptic coupling [1] (it is somewhat controversial). ...[text shortened]... logical_neuron_model

Especially note the last section in [2] - Embodiment in Electronic Hardware.

What is the real goal here, "artificial intelligence" sounds rather nice, but what is it if we see it? As I pointed out earlier once a computer stops doing what it is expected it is no longer useful, that the intelligence of a 14-year-old girl/boy being stubborn, or merely the processing ability of the computer has become toast? Unless there is something there living, I think the faster and more complex you can get the computer upgraded as well as the sophistication of the software it will still be no different than using an abacus at high speeds.

@KellyJay
Not according to these guys:

https://www.newscientist.com/article/mg22329844-000-supercomputers-make-discoveries-that-scientists-cant/

@KellyJay
You clearly haven't been keeping yourself up to speed with modern technology.
There are already clear examples of AIs that are self-taught, i.e do NOT just blindly follow some program their creators made for them, and are now common place.
A good example of that is alphazero;

https://en.wikipedia.org/wiki/AlphaZero

This AI can beat all grandmasters at chess but wasn't told of anything about chess strategy but rather had to learn chess strategy all by itself by playing many games against itself!

AIs are now often designed to have what is called 'intuition', which isn't too dissimilar from human intuition. In fact, one mental task after another that was previously thought to be strictly the preserve of humans, not computers, are now being mastered by AIs and, presumably, it would be just a matter of when, not if, an AI will be able to do all of them and thus do all the mental tasks a human can do (possibly excluding having feelings and emotions but even that might eventually change!).