Altruistic robots produced through evolution

[Triple Zero]

The tierra robots use them, but in the specific, Sigmatic keeps harping on about dedicated ANN chips.

now you're confused?

Tierra is an artificial life simulation, sort of inspired by a "primordial soup" metaphor, using organisms running concurrently as virtual machine language processes. It does not feature ANNs [Artificial Neural Networks], nor robots.

The robots in the OP were controlled by neural nets, although it didn't say what kind.

Once one learns the math behind an ANN and how it learns and does pattern recognition, one might get the same sort of desillusionment as I did, as an ANN--while roughly based on the idea of a brain--it not so much like a brain, as more it is a stochastical method to numerically approximate a transformation from a high dimensional space to a lower dimensional one.

That doesn't mean, however, that there aren't awesome balls things to do and discover in artificial intelligence research, just that the neural network direction is probably not the place. Even if you want to equip one of your AI bots with some sort of basic pattern recognition self-learning module, most of the time there are better [and often simpler] algorithms available than neural nets. It's just that "neural network" is a term that has a lot of appeal, as opposed to Kohonen Self-Organizing Map, Learning Vector Quantization or k-Nearest Neighbour classification.

[Requia ☣]

Er, ok, not the Tierra thing, I thought this was a continuation of the old program.

And yes, I was bitching about the comparison to brains earlier.

[Captain Utopia]

so. how do we make an insane neural network then eh?

Sometimes I fear that AI will always be more breathtakingly sane than humans.  But then again, eradicating a mindless parasite intent on aimlessly consuming all of the resources it can reach, seems an eminently sane decision.

[Phosphatidylserine]

I've seen crazy bots (not crazy ANN based bots, mind you, but other kinds). They tend to be more thought-disorder/mild-schizophrenia crazy, rather than get-the-spiders-out-of-my-brain-or-the-kid-gets-it/serial-killer crazy -- which is to say they act like automated pinealists. That said, who knows what toobparts or ZalGOS would do with a tactile manipulator.

[Jasper]

The reason I brought up analog neural networks is because they appear to be a much more sophisticated version of what the robots in the OP were using.  I linked to an MIT Technology review article that talks about a separate invention that is tangentially related to the OP.  I wanted to remark that combining these technologies would be interesting, but it came across poorly.

[Phosphatidylserine]

ANN typically stands for artificial neural network, not analog neural network. There is a very different term for an analog electronic neuron simulation.

[Jasper]

ANN typically stands for artificial neural network, not analog neural network. There is a very different term for an analog electronic neuron simulation.

I never even mentioned artificial neural networks, jeez.  Did you even click the link?  Do you think it's a rickroll or something?

[Phosphatidylserine]

The reason I brought up analog neural networks

ANN typically stands for artificial neural network, not analog neural network. There is a very different term for an analog electronic neuron simulation.

We were talking about ANNs for the last page, with the implicit understanding that ANN stood for artificial neural network. I was making this explicit, since you brought up analog neural networks (which would of course also be abbreviated ANN, and given the context, implied strongly that you were expanding the acronym differently from the rest of us).

[Jasper]

Okay, okay.  I get that you were talking about artificial neural networks.  Here's what I'm getting at.  The OP article said

"The ability of robots to orientate, escape predators, and even cooperate is particularly remarkable given that they had deliberately simple genotypes directly mapped into the connection weights of neural networks comprising only a few dozen neurons."

While the link I provided says these other guys have produced a chip that has

200,000 neurons linked up by 50 million synaptic connections, the chip is able to mimic the brain's ability to learn more closely than any other machine.

Please tell me you can see why I mentioned this.

[Requia ☣]

I can see why, and as I've said, that chip is *completely useless* no matter how many neurons are on it, ANNs require flexibility to be useful, for an experiment like the one the OP posted this is doubly true, and according to TFA you posted, the chip isn't as flexible as the simulations are.

[Phosphatidylserine]

On top of that, these days people want to actually SAVE their neural nets. If the power goes out or the backup battery goes on that chip, pop goes the brain.

[Triple Zero]

ANN typically stands for artificial neural network, not analog neural network. There is a very different term for an analog electronic neuron simulation.

I never even mentioned artificial neural networks, jeez.  Did you even click the link?  Do you think it's a rickroll or something?

no, it's worse, it's a popular science article without any explanation whatsoever about what the researchers actually DID.

Okay, okay.  I get that you were talking about artificial neural networks.  Here's what I'm getting at.  The OP article said

"The ability of robots to orientate, escape predators, and even cooperate is particularly remarkable given that they had deliberately simple genotypes directly mapped into the connection weights of neural networks comprising only a few dozen neurons."

While the link I provided says these other guys have produced a chip that has

200,000 neurons linked up by 50 million synaptic connections, the chip is able to mimic the brain's ability to learn more closely than any other machine.

Please tell me you can see why I mentioned this.

well yes, I see why you mentioned this.

and I don't like what I see.

because you seem to imply that more neurons = smarter AI.

I tried to explain in my previous post, but you;re really stuck on your concept of what you think are "neural networks".

see, the things that consist of a few dozen neurons in the robot controllers are completely different things than the stuff in the 200,000 neuron simulation.

for starters, apart from some superficial similarities, the neurons these two systems consist of aren't even the same things, so it doesn't really make any sense to compare them.

the robot controller neural nets are the pattern recognition algorithms I somewhat explained in my previous post, which are commonly known as Artificial Neural Networks, abbreviated as ANNs. using the abbreviation for something else is really confusing. [and yes it is a pretty standard abbreviation, just check how many scientific AI symposia there are with ANN in their title, and notice it always means "artificial"]

now the 200,000 neuron thing. while probably really impressive, as far as I've understood these things are only useful from a sort of neuro/bio/physiological point of view. they're basically bruteforce simulating parts of the brain. usually rat brains, it seems. now these things may do some kind of pattern recognition, if you try hard enough, but that's not really what they're made for, and I wouldn't even know if they'd be any better at it than the (mathematically proven) traditional algorithms.

no they are used for completely different kinds of research. think of it more like a dummy. basically they got this simulation of a part of the brain (afaik they got up to a beam of synapses in the cortex by now) which is built to mimic the physiological/biological original as close as possible. yep, that is including differential equations to model the reaction and diffusion of neurotransmitter chemicals. the cool thing about this, is if you want to test it, and/or figure out how it works, you don't need to cut open a rat every time, and most importantly, you can look at it while it's running.

and yes, those things do exhibit aspects of learning. but probably not the kind of reasoning-learning you are thinking of. that's a bit too high level for a beam of synapses in a rat's cortex. actually I wonder if rats can even reason? anyway that;s not the point, the kind of learning it displays is more like a few levels below classic pavlov conditioning. neurons that get excited more often, get stronger pathways and therefore excite even more often, and that's basically the building block for learning.

but that doesn;t make a cool popular science article. and that is why, if you're really serious about this stuff, you should avoid those articles and click through until you find something written in LaTeX, with formulas in it, that gives you the deep rundown of what they actually do instead of what the scientific journalists make of it.

anyway if you want to make a cool AI like what you're aiming at, i'm not saying it can't be done, but I think it's kind of cumbersome to make a simulation that low level. if you want to replicate the biological functions so that you can learn more about our own brains, sure. but if you just need the behaviour, you're wasting resources on accuracy that is unnecessary. it's already a computer, after all. you could jam much more artificial neurons into a chip if you took some liberties and they dont have to function exactly numerically like biological neurons. after all, the intelligence is not in the matter that it's made of, but in the patterns that appear/emerge in that matter, right?

[Jasper]

Thanks for taking to time to write a comprehensive clarification, 000.  I think I need to spend more time reading academic works on the things I profess to be interested in.

[BabylonHoruv]

Also, altruism isn't a matter of game theory, which is how these robots are operating.  Altruism comes from intentionality (and sentimentality), which these robots tend to lack.

How do you know?

[Jasper]

How do I know what?