Print

Please login or register.
1 Hour 1 Day 1 Week 1 Month Forever
Login with username, password and session length

[Egan_BW]

Now I want to see an adversarial technique that lets you paint things so that humans see them as the wrong thing.
"This, officer? Oh it's just a turtle, as you can see."

[Trekkin]

Now I want to see an adversarial technique that lets you paint things so that humans see them as the wrong thing.
"This, officer? Oh it's just a turtle, as you can see."

That would be known as camouflage.

[Egan_BW]

Yes, but current camouflage doesn't exploit subtle inaccuracies in human perception in order to make them believe something patently false.

[Reelya]

Yes, but current camouflage doesn't exploit subtle inaccuracies in human perception in order to make them believe something patently false.

How do you know that is true a true statement though? Isn't that what camouflage does? We just don't see our own optical illusions as "subtle inaccuracies" in that sense, but isn't that a matter of perspective?

Things that fool current classifier AIs don't fool us so we see our vision as superior to theirs, however, if we make a technological system that sees right through camouflage that can fool any human, aren't we really in the same position then? Sure, our vision systems are much more complex and developed than e.g. Google's photo classifier, but that doesn't mean adversarial examples don't exist for humans - e.g. images that fool human vision but wouldn't fool some hypothetical more advanced vision system.

Another thing to remember here is that adversarial example generation techniques are a core part of modern NN vision training. Having a replicable algorithm that fools the current generation of vision systems means that the technology will advance faster.

[Maximum Spin]

That is, you can train a neural network over an infinite sample of these fluke cases, and you will still have an infinite number of adversarial cases where it will fail.

There is, in fact, only a finite (though incredibly large) number of possible adversarial cases to begin with. That said, it's impossible for any realistically feasible algorithm to achieve perfect classification (including human brains) at useful resolutions; this isn't a flaw in the algorithms as such, but a flaw in the thing we're trying to do.

[Reelya]

What has me concerned, however, is to what degree these adversarial cases can be "learned away." It's my opinion that these things aren't just showing issues with the precise tuning of the weights involved in the neural network, but the very architecture of the algorithm in play. That is, you can train a neural network over an infinite sample of these fluke cases, and you will still have an infinite number of adversarial cases where it will fail.

but there are also an equally infinite variety of camouflage patterns that would fool a human. So, in a theoretical sense there can be infinite patterns which fool the system, but that doesn't mean that the system is limited to be worse than humans.

e.g. say you keep feeding adversarial patterns into an NN, growing it larger and larger as you go to deal with all edge cases. Is there any reason to think it's going to plateau at any particular level of ability, and that that level of ability is necessarily worse than a human? Perhaps there is such a theoretical plateau, but the plateau is above the level of human ability. however, there's no theoretical reason to think the plateau exists, nor is there any reason to think it's in any way related to human vision ability.

So, yep, you can keep feeding adversarial examples into a system, and there will always be more adversarial examples you can generate, ad inifinitum. But that doesn't mean that the vision system's abilities are static or limited to be worse than humans, because we lack any real theoretical reason for that: the faulty premise is that humans don't have "adversarial examples", when we clearly do. They're just not perceptible to us because that is the very nature of adversarial examples, so we don't realize we have them.

[Max™]

*cough*
https://cvdazzle.com/

[Doomblade187]

*cough*
https://cvdazzle.com/

Forgot about that, would actually wear some of these.

[Kagus]

I wear patterned gym clothes, to give the illusion that I'm moving.

[Max™]

You fucking win everything for that.

[Loud Whispers]

I wear patterned gym clothes, to give the illusion that I'm moving.

The messiah has returned

[Max™]

Huh, this actually looks promising for some reason: https://arstechnica.com/gadgets/2018/04/the-galaxy-s9s-dex-pad-launches-may-13-for-99-99/

[Parsely]

Huh, this actually looks promising for some reason: https://arstechnica.com/gadgets/2018/04/the-galaxy-s9s-dex-pad-launches-may-13-for-99-99/

Okay now that is a cool idea. Mobile devices are becoming more and more similar to desktop PCs, so this just seems like a logical next step.

[wierd]

The issue is how much is in the docking station, and how much is baked into the device.

This is not a terribly new idea.  Many "Old" (like REALLY REALLY old) docking stations for first generation laptops had actual ISA slots and stuff inside them, so that more powerful desktop grade hardware could be inside the docking station, cut down on weight and thickness of the device, as well as cutting down on power draw.

short trip down memory lane on docking hardware

With a phone, you are limited more and more by what you can reasonably drive for 4 to 6 hours on a 3500mAh LiPO battery. Granted, ARM processors sip on the juice rather than gorging themselves like a gourmand, but the power draw is still a significant design consideration for these devices and the hardware they have inside them, along with total size and weight.

When you start trying to make these devices also be beefy enough to do desktop oriented tasks (which is the whole point of a dock like this), then you start throwing a vastly different set of needs around, and the underpowered hardware of the phone is going to be a problem. 

Now, it is possible that a great deal of fancy hardware could live inside the dock, attached by a lightning port or something. (Lightning has a high bandwidth and supports DMA, which means you could put lots of useful things on it, like a GPU, a hard disk controller, etc, and not bog down the CPU.) If they expose actual bus pins to the dock, the dock could contain an additional processor, and RAM as well.

I guess what I am saying is that rather than something completely new, I am having an intense feeling of deja-vu.  First generation laptops were so underpowered it was hilarious. Compared to an actual desktop system that you can get for about 300$, your cellphone looks like a computer from 20 years ago, just very small.

It will be interesting to see where this goes, but I see it as being a novelty, not a new paradigm, and see it as just a modern evolution of an old idea.

[Sheb]

Why would the power drain be an issue? Presumably the dock would charge the phone, after all you already need a power supply for the monitor.