back to article Photonic processor can classify millions of images faster than you can blink

Engineers at the University of Pennsylvania say they've developed a photonic deep neural network processor capable of analyzing billions of images every second with high accuracy using the power of light. It might sound like science fiction or some optical engineer's fever dream, but that's exactly what researchers at the …

  1. Christoph

    What it will actually get used for

    "the technology has implications for a variety of fields"

    Detect Image - classify as enemy soldier - shoot. All in a fraction of a second.

    The only difficult bit will be the hand-waving cover up of all the false positives.

    1. Lord Elpuss Silver badge

      Re: What it will actually get used for

      If this is US/UK miltech there won't be any false positives to cover up; but they will release a nice report after a few months showing the categories that were identified by their infallible chip.

      1. Legitimate enemy combatants

      2. Legitimate enemy combatants dressed as civilians, but we weren't fooled

      3. Legitimate enemy combatants dressed as civilians, 11,000 miles away from the warzone, in Walmart - but our chip is so good it got the bastard anyway

      4. Legitimate enemy combatants who'd somehow managed to morph into the shape of a goat, car, bridge, traffic light etc, but our chip could see through their sneaky Commie camouflage and still take them out.

      Our chip is perfect, it just found some "Enemy combatants in non-standard scenarios".

      1. Anonymous Coward
        Big Brother

        Re: What it will actually get used for

        It's not the military, it's the police that are going to jump all over this.

        1. stiine Silver badge

          Re: What it will actually get used for

          You do recall that most image processing systems have problems with non-caucasian skin tones. if they adopt it, it will be another spectacular failure.

    2. elsergiovolador Silver badge

      Re: What it will actually get used for

      This is why China built tik-tok. They are using the videos Western users upload every second to build a model that can recognise Western person with near 100% accuracy.

      They'll be using it to control immigrants and also for military drones.

      If the time comes, they'll be able to execute all Westerners in China swiftly and spare their own people.

      1. Shades

        Re: What it will actually get used for

        Tell me, is the earth flat too?

      2. Ian Johnston Silver badge

        Re: What it will actually get used for

        They are using the videos Western users upload every second to build a model that can recognise Western person with near 100% accuracy.

        Hot dog / not hot dog?

        How do they deal with western people of Chinese heritage?

        1. elsergiovolador Silver badge

          Re: What it will actually get used for

          How do they deal with western people of Chinese heritage?

          The AI can pick up difference in diet, behaviour and mannerism, which may not be obvious to the naked eye, but a person living in China will look slightly different to a person of Chinese descent living in the US.

          They have a ton of data to do that, for instance to catch differences based on location of the person, their age, usage patterns.

          It's a brilliant tool - in that when there are some edge cases or some extra data is needed, the virtual fans (CPC workers) can persuade the target to record something - like film themselves walking in the town, eating or making close ups of their face and enrich their models.

      3. Robert Grant

        Re: What it will actually get used for

        The camouflage will be to not do little dances while making different exaggerated facial expressions.

      4. stiine Silver badge

        Re: What it will actually get used for

        What in hell makes you think the Chinese government would stop from killing additional 10-20 million of their own?

        This tech reminds me of the Doonesbury strip during the 1st gulf war, whiere it showed a missle heading for an enemy headquarters building, passing through a window, flying up a stairwell, down a hall, through a conference room full of enemy leaders, out another window, and finally blowing up a hospital.

      5. Arthur the cat Silver badge
        Black Helicopters

        Re: What it will actually get used for

        If the time comes, they'll be able to execute all Westerners in China swiftly and spare their own people.

        I presume you buy your tinfoil in bulk.

    3. pimppetgaeghsr

      Re: What it will actually get used for

      Same as human soliders then.

    4. Eclectic Man Silver badge

      Re: What it will actually get used for

      Maybe they could sell an upgraded version to the 'Auto-pilot' sorry not-quite-self-driving vehicle manufacturers to stop them crashing into, for example, oncoming vehicles, cyclists etc:

      https://www.theregister.com/2022/05/16/automous_car_ada_system_tests/

    5. adam 40 Silver badge
      Paris Hilton

      Re: What it will actually get used for

      Porn

      1. Francis Boyle

        Re: What it will actually get used for

        Reverse cowgirl, 69, missionary (boring), I don't know WTF that is, etc. Could be very useful.

  2. WhereAmI?

    The Chinese Communist Party are going to absolutely luuuvvvv this.

    1. pimppetgaeghsr

      Agreed, there is a massive mounting pressure for all these .ai bubble companies to turn profitable and the VC capital is likely getting impatient. It turns out paying a bunch of postgrads and postdocs to sit around over-fitting models and cleaning data sets isn't a viable business, the only real value is applying it to massive platforms that already have the data such as all the big FAANG platforms.

  3. Big_Boomer Silver badge

    Skynet is your friend

    I'm all for technological advances but far too many these days seem to be for the purposes of intruding into our private lives. Someone needs to explain to the tech companies that we don't want to live in Dystopia, no matter how much money it makes them.

    1. Filippo Silver badge

      Re: Skynet is your friend

      Researchers should probably spend some time thinking about the ethical implications of their work, but decisions on what to do with some new bit of tech are ultimately political - not a job that scientists are generally good at.

      From another point of view: if the only thing that stands between us and dystopia is technical issues... well, then we are already screwed.

  4. Anonymous Coward
    Anonymous Coward

    So it's not a Positronic Ray

    then?

  5. Doctor Syntax Silver badge

    Given that it's a 5x6 pixel image, possibly with only 1 bit per pixel (we're not told otherwise but its only discriminating between two letters, then there's an obvious need for the rider "for some value of image".

  6. Anonymous Coward
    Anonymous Coward

    Yawn!

    > The actual test involved classifying hand-drawn "P" and "d" characters projected onto the chip.

    Come back and talk to me when it can classify the difference between a 1 an l or an I and that's just staying with letters from the US-ASCII character set. Telling the difference between a dog and cat as well as toddler would be slightly more impressive.

    1. Paul Herber Silver badge

      Re: Yawn!

      Peeps on the interwebs need to know difference between cats and KITTENS!!!!!

      1. Sir Awesome

        Re: Yawn!

        No we motherfucking do not - they're all kitties! :D

      2. Arthur the cat Silver badge

        Re: Yawn!

        Peeps on the interwebs need to know difference between cats and KITTENS!!!!!

        Practical experience in that area of research.

    2. Tony W

      Re: Yawn!

      There is very often no consistent visual difference between the images of similar looking alpha-numeric characters. We manage only because the difference is quite often detectable with the aid of a lot of human-understandable context.

      1. Anonymous Coward
        Anonymous Coward

        Re: Yawn!

        That was precisely my point a real image recognition system is going to need to process a lot more than just the content of a jpeg file. That even before getting onto understanding the meaning of an image and making value judgments. Someone below mentions the porn/not porn question, no one can tell you what porn is they just know it when they see it. Other cases are much harder, photos of a baby in a bath, happy parents with happy child or abuse?

  7. Brewster's Angle Grinder Silver badge

    Analogue computing. It's the future!

  8. Howard Sway Silver badge

    The actual test involved classifying hand-drawn "P" and "d" characters projected onto the chip

    This could be useful, as long as you don't accidentally use it udside pown.

    1. milliemoo83

      Re: The actual test involved classifying hand-drawn "P" and "d" characters projected onto the chip

      "This could be useful, as long as you don't accidentally use it udside pown."

      Udsipe pown, Shirley?

      1. Paul Herber Silver badge

        Re: The actual test involved classifying hand-drawn "P" and "d" characters projected onto the chip

        ndsıpә pomu, Shirley.

  9. Paul 195
    Childcatcher

    Dystopia beckons

    "Your scientists were so preoccupied with whether or not they could, they didn’t stop to think if they should."

  10. Forestmania

    Facial recognition

    Facial recognition on an industrial scale, able to discern enough features to establish identities despite masks.

  11. To Mars in Man Bras!
    Thumb Up

    Sounds Like Just What I Need!

    The other day I found a 5GB zip archive which I'd downloaded from Flickr when they did away with their 1TB free storage [I'd used it as a backup]. When expanded I was left with 20.000+ images.

    On my backup hard drive at home, I have an unknown [but possibly even larger] amount of photos, stored in a hierarchical folder structure: YEAR > DATE > EVENT. I need to compare what was in the Flickr archive with what's dispersed throughout the myriad folders in that hard drive archive.

    In their infinite wisdom, Flickr has renamed all the photos when I uploaded them. So I can't employ the usual suspects such as diff or find etc. to see which photos in the Flickr download I already have. I suspect most of them, but don't want to bin any just in case.

    So let me know when this Photonic processor is cheaply available. It could run through my tens of thousands of images in a nanosecond and let me know which ones are duplicates.

    1. mattaw2001

      Re: Sounds Like Just What I Need!

      For your specific problem I would consider the opensource (Linux/Windows) DigiKam: https://www.digikam.org/

      It can fingerprint files and look for matching ones for you to eliminate so you can add the images to your existing library (backup first!), fingerprint them all and show you the matches sorted by 100%% -> 0% and eliminate the duplicates.

      This is the only thing that works for me as I used google photos and they mess with the metadata so you can't use file hashes to find duplicate images anymore, sadly.

      1. To Mars in Man Bras!
        Thumb Up

        Re: Sounds Like Just What I Need!

        Thanks. I'll look into it

  12. Pascal Monett Silver badge

    30 pixels

    That might be enough to recognize a toenail.

    My Samsung A3 takes pictures that are 4128 x 3096. That makes for 12 780 288 pixels, which is 426 000 times their capacity of treatment.

    I think this tech needs a bit of upgrading before we can worry about Big Brother recognizing our faces.

  13. LeftyX

    So it's a perceptron

    This sounds like an all-optical version of the perceptron, invented in 1943 (!), but of course much smaller.

  14. Caver_Dave Silver badge

    Will social media get behind it?

    If scaled up then the killer app is surely identifying unsavoury content being uploaded to social media, rather than having staff look at it once it has been flaged by the public.

  15. Eclectic Man Silver badge
    Joke

    Simulating brain neurons

    "The standalone light-driven chip – this isn't another PCIe accelerator or coprocessor – handles data by simulating brain neurons"

    BUT, can it remember what it went upstairs for when it gets there? If so, please may I have one, or two, yes two, please, in case I forget where I put the first one...

  16. An_Old_Dog Silver badge
    Paris Hilton

    Porn/Not-Porn Classification

    Computer: "Not porn, next ... not porn, next ... porn, next ... porn, nex ... wait, wait, go back! Oh, yeah ..."

  17. JDPower666

    All the armchair critics slagging off the image size and ultra basic abilities seem to forget, the first microprocessors were completely useless in reality, and look where we are now. It's easy to be critical whilst other people are out there making new and innovative technologies a reality.

    1. Doctor Syntax Silver badge

      Over many years I've seen all sorts of technologies show amazing promise and disappear. With this one it's relevant to ask how the difficulty scales with number of pixels, bits per pixel and object categories and with object complexity.

  18. Pen-y-gors

    What's a "category"?

    "the photonics chip was able to categorize an image in under 570 picoseconds with an accuracy of 89.8-93.8 percent."

    For a given value of 'categorise'?

    a) There's a building in the picture

    b) It's a picture of 17 Railway Sidings, East Cheam

    a) There's a human in the picture

    b) It's a picture of a soldier in the Russian 17th Motorised Artillery Brigade, who isn't a POW.

    1. Eclectic Man Silver badge

      Re: What's a "category"?

      I wonder how accurately a human would manage the categorisation. Obviously not nearly as quickly, but the ones the chip mis-categorised or simply gave up on would be interesting. I speak as someone who sometimes has difficult reading my own handwriting. (It has got a lot worse since everything had to be typed on a computer keyboard instead of written down longhand.)

      Also, upvote for East Cheam, fictional home of Tony Hancock (although I thought he lived at 17 Railway Cuttings, but then my memory isn't what it used to be).

      1. Doctor Syntax Silver badge

        Re: What's a "category"?

        It needs to be able to discriminate between an image of 17 Railway Cuttings and 17 Railway Sidings.

        1. A.A.Hamilton

          Re: What's a "category"?

          "... discriminate between an image of 17 Railway Cuttings and 17 Railway Sidings."

          Using 54 year-old traces of humour (at genius level), melancholia, alcoholism, self-loathing, lovability, sensitivity, as key discriminating factors,I assume?

        2. Anonymous Coward
          Anonymous Coward

          Re: What's a "category"?

          I met a relative of Tony Hancock in the pub the other day... he was running through Half Hour quotes like a US student quoting Monty Python until I expressed my mild annoyance by grinding my first in my palm. It was only later that I realised my behaviour was rude... Really I should have offered him a punch up the bracket.

  19. Spherical Cow Silver badge

    Captcha

    Get back to me when it can select all images with a fire hydrant.

  20. RLWatkins

    Useful, but let's don't work ourselves into a frenzy

    If this is what I think it is, an optical version of a neural net built from a block of glass fibres, it has to be hard-wired.

    That means that it has to be hard-wired for a specific purpose. At low levels this isn't all that bad: we have hardwired edge and area detection for our own eyes, providing additional information derived from a field of pixels which is useful for recognizing what we see.

    Scaling that up won't be one of those square-of-the-number-of-pixels things since that sort of recognition depends on nearby pixels rather than correlating information from distant parts of the visual field.

    So far, so good.

    The problem comes with hard-wiring the higher levels of recognition.

    You want to recognize numbers and letters? Roman letters? Cyrillic? Arabic? Telugu?

    You want to recognize faces? Whose faces? Maybe worth hard-wiring to recognize specific features, but beyond that not so much.

    So at some point someone has either to create a *programmable* optical neural net to sit atop this device, or to use the device to enhance the input to more conventional software.

    Sounds useful, but this isn't the be-all and end-all of image recognition. Recognize it for what it is, to wit a useful step forward, but let's don't get all dizzy over it.

POST COMMENT House rules

Not a member of The Register? Create a new account here.

  • Enter your comment

  • Add an icon

Anonymous cowards cannot choose their icon

Other stories you might like