Everyday Objects Can Run Artificial Intelligence Software (science.org) 51
Slashdot reader sciencehabit quotes Science magazine: Imagine using any object around you—a frying pan, a glass paperweight—as the central processor in a neural network, a type of artificial intelligence that loosely mimics the brain to perform complex tasks. That's the promise of new research that, in theory, could be used to recognize images or speech faster and more efficiently than computer programs that rely on silicon microchips.
To demonstrate the concept, the researchers built neural networks in three types of physical systems, which each contained up to five processing layers. In each layer of a mechanical system, they used a speaker to vibrate a small metal plate and recorded its output using a microphone. In an optical system, they passed light through crystals. And in an analog-electronic system, they ran current through tiny circuits.
In each case, the researchers encoded input data, such as unlabeled images, in sound, light, or voltage. For each processing layer, they also encoded numerical parameters telling the physical system how to manipulate the data. To train the system, they adjusted the parameters to reduce errors between the system's predicted image labels and the actual labels.
In one task, they trained the systems, which they call physical neural networks (PNNs), to recognize handwritten digits. In another, the PNNs recognized seven vowel sounds. Accuracy on these tasks ranged from 87% to 97%, they report in this week's issue of Nature. In the future, researchers might tune a system not by digitally tweaking its input parameters, but by adjusting the physical objects—warping the metal plate, say.
The team is most excited about PNNs' potential as smart sensors that can perform computation on the fly. A microscope's optics might help detect cancerous cells before the light even hits a digital sensor, or a smartphone's microphone membrane might listen for wake words. These "are applications in which you really don't think about them as performing a machine-learning computation," they say, but instead as being "functional machines."
To demonstrate the concept, the researchers built neural networks in three types of physical systems, which each contained up to five processing layers. In each layer of a mechanical system, they used a speaker to vibrate a small metal plate and recorded its output using a microphone. In an optical system, they passed light through crystals. And in an analog-electronic system, they ran current through tiny circuits.
In each case, the researchers encoded input data, such as unlabeled images, in sound, light, or voltage. For each processing layer, they also encoded numerical parameters telling the physical system how to manipulate the data. To train the system, they adjusted the parameters to reduce errors between the system's predicted image labels and the actual labels.
In one task, they trained the systems, which they call physical neural networks (PNNs), to recognize handwritten digits. In another, the PNNs recognized seven vowel sounds. Accuracy on these tasks ranged from 87% to 97%, they report in this week's issue of Nature. In the future, researchers might tune a system not by digitally tweaking its input parameters, but by adjusting the physical objects—warping the metal plate, say.
The team is most excited about PNNs' potential as smart sensors that can perform computation on the fly. A microscope's optics might help detect cancerous cells before the light even hits a digital sensor, or a smartphone's microphone membrane might listen for wake words. These "are applications in which you really don't think about them as performing a machine-learning computation," they say, but instead as being "functional machines."
Frying Pan? (Score:2)
Finally an A.I. that can cook me dinner. (Score:2)
Yes, imagine. If AI only was AI (Score:1, Informative)
Until we understand how learning works, really understand it, not fake it, we will have lots of crappy AIs that behave just like Clippy, that can help with some limited range of things, but give stupid answers outside of those tiny limits.
Until we understand how learning works, IGNORE ALL AI HYPE.
Re: (Score:2)
This is more properly described as a neural net.
Neural nets can recognize patterns once trained.
Very clever what they have demonstrated.
Re: (Score:2)
This is more properly described as a neural net.
Neural nets can recognize patterns once trained.
Very clever what they have demonstrated.
It's literally in the description:
the researchers built neural networks
And later on:
In one task, they trained the systems, which they call physical neural networks (PNNs),
Re: (Score:2)
You mean... Artificial Stupidity? Yeah, we're great at doing that!
"One day, machines will exceed human intelligence." - Ray Kurzweil
"Only if we meet them half-way." - Dave Snowden
Looks like we're really doing our "Idiocracy" best to meet them half-way.
Re: (Score:2)
Re: (Score:2)
Lots of fast, dumb calculations is not *and never will be* AI.
Yet, that is also how biological neural networks function. The main difference is that the activation function is analog, but the principle is the same.
Re: (Score:2)
Re: (Score:1)
And, given the faster and faster and faster and still not learning computers, and the main difference that real intelligence really learns and fake ones do not,
I contend that you are wrong, and you and your colleagues can keep throwing cycles at your solution, but will never solve the problem.
Re: (Score:2)
Re: Yes, imagine. If AI only was AI (Score:2)
It would be more like "Wright brothers create artificial life!! Today, Orville and Wilbur Wright demonstrated a mechanical bird and predicted that soon flocks of these creatures would roam the skies and could be harnessed like horses to do tasks for man!"
Re: (Score:1)
Re:Yes, yes, but (Score:2)
But should they? (Score:2)
I have no problem with new technology, I love technology. The problem I have is with how companies are choosing to use technology, which is specifically not to the benefit of the buyer. I know almost everyone is willing to trade away their privacy for convenience but Faustian bargain at best. I certain that integrating technology into otherwise basic tools will inevitably lead to further decay in privacy.
So this all piles up to one question: is it in the long-term interest of our society to do away with
Re: (Score:2)
Re: (Score:2)
doesn't need to report to the cloud everything it does
Neither do SmartTVs, or "fitness trackers" and yet they do.
Re: (Score:2)
is it in the long-term interest of our society to do away with every last shred of privacy?
If we could actually do that, such that literally everyone's privacy would be eliminated, then it would arguably make the world a better place. The fundamental problem with such an idea is that it never actually works that way. There are always haves and have-nots, in this case those who have privacy and those who do not. You wind up with politicians looking into people's bedrooms, without those people being able to look back into those politicians' offices, or into corporate boardrooms.
Privacy has to be gi
Re: (Score:2)
Privacy has to be given to everyone or no one for it to be fair.
At this point, you only need for people to reject products that spy on them.
Re: (Score:2)
At this point, you only need for people to reject products that spy on them.
So don't use the internet?
Re: (Score:2)
The internet is no particular product, so no. You definitely should reject some websites though. I browse using a cookie whitelist and many privacy enabling extensions.
Neurons (Score:2)
Pretty soon we'll be doing this with something far more powerful: actual human neurons! It'll be just like this experiment: we'll tweak the input parameters via the ingesting of drugs like SSRIs, hormones, neuroactive pharmaceuticals, or even repeated auditory and optical stimuli, and the output can be anything: observed behavior, what you say on social media, your accent, your beliefs, and any way you already communicate your thoughts. This whole experiment can be as simple as showing the optical neurons i
Re: (Score:2)
Re: (Score:2)
Somebody clearly should reduce their drug intake (Score:3)
Because this can pretty much only be the result of bingeing on psychoactives.
Re: (Score:2)
If you liked this story, you might also enjoy the science fiction of Rudy Rucker. Heavy on psychedelics, math, and AI.
Re: (Score:2)
If you liked this story, you might also enjoy the science fiction of Rudy Rucker. Heavy on psychedelics, math, and AI.
Thanks, I have read some of his stuff.
Ridiculous "science journalism" interpretation. (Score:1)
Boiled down to its essentials, researchers used mechanical, optical, and electronic systems as analogue computers to perform neural network computations.
It has zero to do with using a frying pan or a glass paperweight as the central processor in a neural network.
These experiments are in fact quite elaborate, don't use "everyday objects," and involve a non-negligible amount of energy to set the input parameters in their physical systems. In my opinion it is not quite clear that they can outperform digital ne
Oh look, we're making an analog computer... (Score:2)
Oh look, we're making an analog computer using unusual components and calling it a breakthrough. Analog computers are centuries old, and even analog neural networks are decades old.
Re: Oh look, we're making an analog computer... (Score:2)
Re: (Score:3, Insightful)
Glad I didn't have to scroll down too far to find this kind of push-back. Call me when my frying pan can implement NAND and has decent fan-out.
I don't know if we should blame shoddy science journalism here, shoddy publish-or-perish academic "more of the same" to earn a "piled higher and deeper", or what; but just because a canyon can make an echo effect obviously doesn't mean it can do the same thing as a general purpose computer that was programmed to make echo effects.
Re: (Score:2)
Exactly.
What a laughable article: "What we do with a digital
computer, we can also do with an analog computer!"
I guess most millennials have no idea what an analog
computer it.
A tech demo with very little real-world value? (Score:1)
1. Integrating any kind of processing into the object itself means you have to replace the whole object if any part of the 'smart property' fails. With a separate sensor, maintenance is possible.
2. Sounds like it needs an outside power source.
3. Actually communicating the result to the outside word needs more components/power/connections, which seems pretty much to negate the benefit of having the processing itself embebbed so deeply into an item.
Let
Re: (Score:3)
It's a fundamentally dumb idea because you want the processing to be done on a device you can and probably will upgrade, where you can therefore upgrade the processing ability. Peripherals are best fairly dumb, with just enough intelligence to provide some kind of standard interface so that drivers don't become a problem.
Re: (Score:2)
Re: (Score:2)
Wrong (Score:1)
OW OW OW MY EYES MY EYES (Score:2)
Re: (Score:2)
Maybe even ... (Score:2)
inevitable thundering doom (Score:3)
missing the why (Score:2)
I'm missing the part about why you'd want an AI frying pan. I do not agree that 'put a chip in it' applies to everything. We all know how this ends, the frying pan reports what I'm cooking back to some shady marketing company and I get constant phone calls about renewing my frying pan warranty. Just imagine how much more advertising you could be subjected to in your life if everything around you, and at all times, reported on you and made suggestions to you.
Re: (Score:2)
Mechanical Computers (Score:3)
Re: (Score:1)
The 1950s called... (Score:2)
Re: (Score:2)
Something I wanted to do for a science museum (Score:2)
This kind of physical machine learning model was something I always wanted to try making for a local science museum. The idea was to create a 32x32 grid of boxes and allow kids to fill the boxes with small BBs in the general shape of a letter of number, then let the BBs get siphoned through a series of physical gates whose weights were adjusted based on the MNIST data set. The bottom of the machine would be 36 buckets representing each letter and single digit. The BBs would then pool into the buckets and h
Goldberg (Score:2)
A Goldberg "AI." What could possibly go wrong?