Bay 12 Games Forum

Please login or register.

Login with username, password and session length
Advanced search  
Pages: 1 ... 73 74 [75] 76 77 ... 144

Author Topic: Tech News. Automation, Engineering, Environment Etc  (Read 85873 times)

Reelya

  • Bay Watcher
    • View Profile
Re: Tech News. Automation, Engineering, Environment Etc
« Reply #1110 on: November 08, 2017, 02:08:47 pm »

There's always the awful solution.

Everyone wants to be the one person whose life all self-driving cars are designed to spare at all costs, whether they're in them or not; this is no different from what they want from human-driven cars, but people are more comfortable making impossible demands of robots.

Well, then, why not sell that? (That question is rhetorical.) Auction off avoidance priorities, and just have the cars communicate with each other and with the pedestrians' phones to decide who to hit based on a greedy algorithm that steers them away from the highest-priority individuals in the vicinity.

Yes, it's awful in countless ways. But it makes the automated car executives more money than any other solution I've seen so far, so it's probably going to be the one we end up implementing.

you're probably right about that actually, I'd sort of missed this post before, it ties into what I've written too. Some ways this trade-off could be monetized include building it into your insurance premiums. e.g. if you get a "greedy" algorithm such as "steeravoid 3000" which always maximizes your chance of life at the expense of others, then the insurance companies are going to start taking note of which algorithm you've chosen to implement. And then they're going to gouge you on third-party insurance. It would basically be similar for all forms of damage and potential medical expenses, the people who can afford the least will end up with AI algorithms that minimize potential costs for all insured road-users, whereas the rich won't care - they'll happily pay extra insurance to be covered for having an AI that is willing to sacrifice others for their own safety. But it won't be phrased as "willing to run down a woman with a baby, if it protects the car occupant", it will be phrased as "maximize safety at all costs". This AI you bought that "maximizes safety at all costs" might then even make some terrible decisions that the real driver wouldn't have - like running over 20 school children instead of increasing risk for the car occupants.

Car AI is going to be important in cases such as accident forensics, so I imagine a TON of laws are going to be passed so that the choice of AI you have is something you need to tell people about. It's not like linux vs windows on your PC - if you change the code on a life or death machine that can kill other people, the government and insurance agencies will want that information, and they'll legislate to get it. e.g. if they find out you changed your self-driving algorithm to a "greedy" algorithm to avoid damage, and it smashed someone else up, good luck with insurance payouts/premiums. And with full computer logs from both robot cars, if you ran someone off the road to protect yourself, they're going to notice that and make you pay for it.

So there should evolve an eco-system around this where car-makers, AI designers and insurance companies, police and government all play off each other here.
« Last Edit: November 08, 2017, 02:29:01 pm by Reelya »
Logged

Bumber

  • Bay Watcher
  • REMOVE KOBOLD
    • View Profile
Re: Tech News. Automation, Engineering, Environment Etc
« Reply #1111 on: November 08, 2017, 02:10:01 pm »

With any luck that idea would be immediately struck down in courts. This is essentially a "pay-to-not-die" scenario; it violates basic human rights. Yes, it's true we generally have to pay for food, shelter, and other basic necessities, but if we are unable to do so, our rights state that we are provided them nonetheless.
Risks are different from necessities. Living in a well-funded, low-crime area isn't a right, even though it affects your chances of dying. Death panels!
Logged
Reading his name would trigger it. Thinking of him would trigger it. No other circumstances would trigger it- it was strictly related to the concept of Bill Clinton entering the conscious mind.

THE xTROLL FUR SOCKx RUSE WAS A........... DISTACTION        the carp HAVE the wagon

A wizard has turned you into a wagon. Was this inevitable (Y/y)?

syvarris

  • Bay Watcher
  • UNICORNPEGASUSKITTEN
    • View Profile
Re: Tech News. Automation, Engineering, Environment Etc
« Reply #1112 on: November 08, 2017, 02:23:44 pm »

Quote from: Ree
So new situations are often phrased as being variants of the Trolley Problem so that we can make use off 50 years of ethics research rather than going in blind to every new situation, which wouldn't allow us to tap into 50 years of insights and research on the existing framework.

It's cute that you think the other people here care to actually educate themselves on the subject.  Yeah, there's plenty of information out there, but you don't need that information to form an opinion and then defend it to the death.

Reelya

  • Bay Watcher
    • View Profile
Re: Tech News. Automation, Engineering, Environment Etc
« Reply #1113 on: November 08, 2017, 02:32:59 pm »

Nah, I wasn't expecting them to do that, just pointing out how the specifics that are being discussed were in fact from a 50-year-old academic thought experiment. The TP is analogous to a lot of real-world decisions, but it's a clear mistake to start applying it literally to every new situation, as if every new ethical situation was only analogous to the TP if it involved 5 pedestrians being run over. If we go that route, then something that's supposed to be a tool that lets us abstract ethical problems and decide if they're really comparable or not, gets reduced to discussing inane details instead.
« Last Edit: November 08, 2017, 02:39:47 pm by Reelya »
Logged

Sheb

  • Bay Watcher
  • You Are An Avatar
    • View Profile
Re: Tech News. Automation, Engineering, Environment Etc
« Reply #1114 on: November 08, 2017, 02:44:36 pm »

Yeah, there's plenty of information out there, but you don't need that information to form an opinion and then defend it to the death.

Can I sig this?
Logged

Quote from: Paul-Henry Spaak
Europe consists only of small countries, some of which know it and some of which donít yet.

Reelya

  • Bay Watcher
    • View Profile
Re: Tech News. Automation, Engineering, Environment Etc
« Reply #1115 on: November 08, 2017, 03:15:05 pm »

Trekkin

  • Bay Watcher
    • View Profile
Re: Tech News. Automation, Engineering, Environment Etc
« Reply #1116 on: November 08, 2017, 03:19:12 pm »

Can I sig this?

Sigged!

You sigged him asking if he can sig someone else's post?
Logged

scourge728

  • Bay Watcher
    • View Profile
Re: Tech News. Automation, Engineering, Environment Etc
« Reply #1118 on: November 08, 2017, 06:50:57 pm »

No, Sheb, I refuse to be sigged yet again.  I think there's three people running around with me in their sigs, I don't think I could take the stress of a fourth.  My skin crawls just imagining it...
sigged

Jopax

  • Bay Watcher
  • Cat on a hat
    • View Profile
Re: Tech News. Automation, Engineering, Environment Etc
« Reply #1119 on: November 09, 2017, 05:30:56 pm »

AI learning cuts down render time from dozens of hours to minutes.

Doubtful if they'll ever release it for wider use since the commercial application is pretty damn huge in terms of time saved but it's still a damn cool thing to see happen.
Logged
"my batteries are low and it's getting dark"
AS - IG

McTraveller

  • Bay Watcher
    • View Profile
Re: Tech News. Automation, Engineering, Environment Etc
« Reply #1120 on: November 09, 2017, 07:55:33 pm »

AI learning cuts down render time from dozens of hours to minutes.

Doubtful if they'll ever release it for wider use since the commercial application is pretty damn huge in terms of time saved but it's still a damn cool thing to see happen.
Ersatz clouds!  Awesome!

It's actually... interesting.  I wonder if it could somehow apply to all those discussions about people wondering about the computational power required to "compute" the universe. I've heard some people hand-wave it as "the universe isn't computed, it just evolves" but perhaps something like this AI stuff applies.  For instance, if all you had was the output of the ANN-produced clouds, would you come up with the physics behind cloud scattering?  Or does it only work the other way - where you have to already know how a cloud works to make a neural network that can make things that look like clouds, but aren't really clouds.
Logged

Trekkin

  • Bay Watcher
    • View Profile
Re: Tech News. Automation, Engineering, Environment Etc
« Reply #1121 on: November 09, 2017, 08:12:04 pm »


It's actually... interesting.  I wonder if it could somehow apply to all those discussions about people wondering about the computational power required to "compute" the universe. I've heard some people hand-wave it as "the universe isn't computed, it just evolves" but perhaps something like this AI stuff applies.  For instance, if all you had was the output of the ANN-produced clouds, would you come up with the physics behind cloud scattering?  Or does it only work the other way - where you have to already know how a cloud works to make a neural network that can make things that look like clouds, but aren't really clouds.

In brief (and knowing full well that I'm opening myself to being barraged by a million poorly written gee-whiz-robots articles from Wired or wherever), neural networks are excellent classifiers in part because they're efficient at pruning down huge state spaces in ways that aren't immediately apparent to human programmers -- which also means that it's not necessarily possible to learn anything transferable from the final network topology. You don't need to know how a cloud works, but you do need to know how to write code that can stochastically produce clouds.
Logged

Ispil

  • Bay Watcher
    • View Profile
Re: Tech News. Automation, Engineering, Environment Etc
« Reply #1122 on: November 09, 2017, 09:18:28 pm »

Those clouds aren't very diverse, though. They all are kind of the same moderately-poofy cloud. You'd need a vastly larger sample to for all varieties of clouds, and I imagine that the neural network system wouldn't be able to really pick up on the firm distinctions in those cloud groups and merge them into not-very-cloudy things.

So it isn't that that AI is rendering clouds. It's that it's rendering a very specific kind of cloud.


Though that isn't to say that's the fault of this specific algorithm. It just seems to be the fault of neural networks in general- they work really well, but leave you with a very small result sample. Though that's not to say that that's not incredibly useful; it's just not as wide a net as first appears.
Logged
"Those who can make you believe absurdities can make you commit atrocities." - Voltaire

"I have never made but one prayer to God, a very short one: 'O Lord, make my enemies ridiculous.' And God granted it."- Voltaire

I transcribe things, too.

Reelya

  • Bay Watcher
    • View Profile
Re: Tech News. Automation, Engineering, Environment Etc
« Reply #1123 on: November 10, 2017, 09:39:29 pm »

Well we're not actually pushing the limits with neural network topologies, are we? So you'd expect to have limited outcomes.

Pretty much everyone even "deep learning" people are in fact just using one-way signal processing which feeds one layer into the next, until it reaches an "output" layer. And then you have some "training" process that alters the weights, but this "training" process is applied external to the network itself. We haven't in fact worked out how to create a "learning" system which is itself part of the normal operation of a neural network.

Consider that real-life neural networks don't even have "layers" and that they have loopbacks, memory and a concept of signals flowing through the network over time, plus they are truly self-teaching: there is no need for a training harness which is outside the network. None of those things are true of the current "deep learning" stuff.

Right now the emphasis is just on adding more layers, more processing power, more training data to get the most out of our one-way layered networks. "deep learning" is just a buzzword that means they have more layers, thus they need more processing power. The network topology isn't actually any less simplistic. There's nothing "deep" in terms of "more complex networking" going on. At some point that's clearly going to come up against some sort of wall where merely adding more processing power and bigger training sets isn't cost-effective vs coming up with smarter topologies.

The types of neural networks we have now are equivalent to having a pocket calculator and not knowing how to use any buttons except "+", "-", "=" then blaming the calculator when it's difficult to calculate a multiplication or division. We're just not exploring a whole lot of the design-space of neural networks at all.
« Last Edit: November 10, 2017, 09:50:07 pm by Reelya »
Logged

bloop_bleep

  • Bay Watcher
  • U want sum seed?
    • View Profile
Re: Tech News. Automation, Engineering, Environment Etc
« Reply #1124 on: November 11, 2017, 02:11:57 am »

Well we're not actually pushing the limits with neural network topologies, are we? So you'd expect to have limited outcomes.

Pretty much everyone even "deep learning" people are in fact just using one-way signal processing which feeds one layer into the next, until it reaches an "output" layer. And then you have some "training" process that alters the weights, but this "training" process is applied external to the network itself. We haven't in fact worked out how to create a "learning" system which is itself part of the normal operation of a neural network.

Consider that real-life neural networks don't even have "layers" and that they have loopbacks, memory and a concept of signals flowing through the network over time, plus they are truly self-teaching: there is no need for a training harness which is outside the network. None of those things are true of the current "deep learning" stuff.

Right now the emphasis is just on adding more layers, more processing power, more training data to get the most out of our one-way layered networks. "deep learning" is just a buzzword that means they have more layers, thus they need more processing power. The network topology isn't actually any less simplistic. There's nothing "deep" in terms of "more complex networking" going on. At some point that's clearly going to come up against some sort of wall where merely adding more processing power and bigger training sets isn't cost-effective vs coming up with smarter topologies.

The types of neural networks we have now are equivalent to having a pocket calculator and not knowing how to use any buttons except "+", "-", "=" then blaming the calculator when it's difficult to calculate a multiplication or division. We're just not exploring a whole lot of the design-space of neural networks at all.

It really, really isn't as simple as you say. For one thing, there are many variations of neural networks. For example, you can build a neural network that takes as input the previous character and outputs the next character. You can then feed it a bunch of text to train it -- due to its design, instead of learning about the text as a whole (which might be very difficult to analyze), it'll learn which combinations of characters are common (so if it sees a 'q', it'll "know" that the next character is more likely to be an 'u' than a 'z'.) This is the mechanism behind Cleverbot and other such chatbots. If Cleverbot was implemented with a neural network that takes the text as a whole instead of word by word it would make much less sense.

If deep learning was as simple as linking up more and more layers as you describe, research in the area would be much more dormant.
Logged
bloop_bleep's Moving Fortress Parts Mod!

Quote from: KittyTac
The closest thing Bay12 has to a flamewar is an argument over philosophy that slowly transitioned to an argument about quantum mechanics.
Quote from: thefriendlyhacker
The trick is to only make predictions semi-seriously.  That way, I don't have a 98% failure rate. I have a 98% sarcasm rate.
Pages: 1 ... 73 74 [75] 76 77 ... 144