back to article Googlers devise DeViSE: A thing-recognising FRANKENBRAIN

You'd think guesswork and advanced science would be natural enemies, but not at Google where a crack team of researchers are trying to mate the two together. In a paper presented on Monday at an artificial-intelligence conference in California, seven Google researchers outlined their image classifier, software that labels …


This topic is closed for new posts.
  1. poopypants

    The real inspector gadget?

    If they hooked this up to London's CCTV cameras, would it be able to recognise criminal activity?

    1. Frankee Llonnygog

      Re: would it be able to recognise criminal activity?

      Yes - if it consists of a bounder in a top hat tying a maiden to the railway tracks and twirling his moustaches. Committing the same crime while dressed as a pedal-bin will baffle even Inspector Google

    2. Frank Rysanek

      Re: The real inspector gadget? (classification of live video)

      At the moment the system could probably recognize some objects in a static way, looking at the video stream as a slide show of static images.

      A proper implementation of "recognizing a crime just happening" would require the system to recognize and classify motions / actions happening in time, in a video stream, preferably in real time. Probably not implemented yet. It would be the next level, for performance reasons if nothing else. A very logical next step...

      Makes me wonder how much information this image classification system can extract from a photo. Break down the photo into a collection or a tree of objects: there's a street with some trees, people on the sidewalk, some cars, and there's a guy swinging a baseball bat (note: try that as a query to Google Images). A human brain would automatically pop out the eyeballs: what? Right there on a street? What or whom is he targetting? Does it look like agression? ... there are a lot of inherent defensive reflexes and experience-based context and attention to motion in a human brain, and emotional aspects, which a relatively spartan neural network trained for automated classification of static images may not possess. Not right now, at least...

      1. druck Silver badge

        Re: The real inspector gadget? (classification of live video)

        What would it make of this video?

    3. TheOtherHobbes

      Re: The real inspector gadget?

      Only if it's someone wielding a shark with a frickin laser beam. Not a nuclear submarine.

      Is anyone else reassured that the world's biggest search engine knows the difference?

  2. Blofeld's Cat


    Am I alone in feeling uneasy that the researchers are training it to recognise sharks?

    It's the red-lined cape with the top hat containing a frikin laser.

  3. Frank Rysanek

    Will spammers be able to manipulate that?

    Now... when this becomes airborne for production service in the Google Images back end, would it be possible to google-bomb this to return pr0n images to some harmless queries?

    Or, maybe Google could use it to *detect* such google bombing attempts :-)

  4. Anonymous Coward
    Anonymous Coward

    And now it tells us...

    ... that 90% of pictures are of a "Pinterest"

  5. frank ly

    I for one am happy about this

    "... distinguishing between categories like dogs, cats, and lawnmowers, ..."

    This is a vital skill in the modern world.

    1. Frankee Llonnygog

      Re: This is a vital skill in the modern world.

      Also a vital skill in Gary Larson cartoons

  6. Frank Rysanek

    How long till consciousness

    There's a growing body of research and knowledge on internal brain functioning and organization: composition of cortical columns, the various neuron species in a biological brain, a coarse global wiring schematic, knowledge of specialized subsystems, knowledge that in some areas the columns "switch purposes", plasticity of the brain, influences from the physical level (various firing/detection tresholds influenced by levels of chemicals, diseases etc), control of and feedback from endocrine glands.

    In terms of computer-based modeling, some scientific teams with origins in biology and neuro-medicine approach this by trying to simulate the biological brain as precisely as possible = computationally simulate the transfer functions and behaviors of the neurons at maximum level of detail, as it is recognized that the pesky low-level details *do* matter, do have an impact on overall brain functioning at macro perespective.

    Other teams (with a knowledge-engineering angle) seem to be more focused on computational performance and cunning topology (with cognitive functionality in mind), taking some inspiration from biology (the introduction of spiking neurons a few years ago) but not necessarily wasting effort on "maximum-fidelity" modeling of the biological brain...

    Google has taught its neural network to classify objects based on their visual and linguistic descriptions combined. It's a neural network, not an old-school AI search term classification engine (which was essentially a network database). This artificial neural network has an inherent neural-style memory with links to external data and BLOB storage, it can classify fresh input data and can retrieve search results based on queries...

    The neural network does not yet have a "flow of thought", a sense of goal or purpose to actively follow, a will, or even a possibility to take autonomous action. Or so I hope...

    Makes me wonder if it would be possible to implement something resembling "flow of thought" without an active will / survival instinct or some such. There is a rudimentary neural engine, capable of sorting and searching visual+linguistic objects and concepts. Perhaps abstract concepts are not so far away. Next, implement associative search capability in that long-term "neural object storage" (maybe it's already there), add some short-term memory (for "immediate attention point"), maybe a filter of some sort (able to limit the "focus" to an object or area) and chain them in a feedback loop. Suppose the "immediate attention" is on a particular object. The associative memory offers a handful of associations, of which the filter/combiner picks a particular area/concept/object. This gets fed back into the "immediate attention" cell. Flow of thought anyone?

    Makes me wonder if this would work without sensory input. Maybe add some relevant input channels to the "filter" stage in the loop (call it a "combiner", op-amp style). Or turn it inside out, and consider it a Kahlman filter made of neural building blocks... Not sure about the purpose or use of this arrangement. Perhaps to extract a model of the sensed reality in terms of objects and concepts, and suggest relevant "mental associations" and possible future developments of the current situation? A mind is probably much more complicated than that...

    Perhaps a simple "flow of thought demonstrator" could be built with much less computational power and "inter-neuron communications bandwidth" than traditionally quoted for a human-level brain. If some biological baggage got "optimized away", perhaps some interesting functionality could be reached "cheaper".

    Scary thoughts. A terminator face fits the topic even better than black helicopters.

    1. RyokuMas

      Re: How long till consciousness

      Tell me about it - always figured Skynet would be born from a search engine gone postal rather than a military system...

      ... and if it's Google, it probably already has behaviourally targetted data to kill every one of us in the way we each fear most. Probably while trying to advertise chainsaws, fire axes and halon to us...

      1. Frank Rysanek

        Re: yes I do feel targetted

        Yesterday I went googling for some milling cutter tools for a hand-held woodwork router. In my mother language, which is not English. Guess what - as I was typing my earlier litany on AI at El Reg, the Google Ad bar on top of the page was flashing some hobby cutter sets at me - pointing me to e-shops in my country. Later yesterday I went googling for a somewhat specific sleeping mattress. Guess what the ad bar shows now... Well at least it's not showing my own employer's ads anymore (which used to be the case for the last half a year).

    2. packrat


      intelligence is definable.

      recognition, manipulation synthesis. And a thermostat does most of that, too.

      sapient and sentient are slightly different

      IQ tests recognize political, academic and creative skills. (abilities) already...

      BUT motor, intellectual and social skills get a higher pirority.

      These NINE words won't help with a few undefined abstracts...

      tactics, strategy, and a few others. (yes, i have defs, BLEAH!)

      where is your problem?


  7. John Smith 19 Gold badge

    Will do untill they can lay hold of enough "focussed" individuals

    to take care of all the hard AI stuff.

    Mines the one with a copy of "A fire on the deep" in the pocket.

This topic is closed for new posts.

Other stories you might like