back to article This is your four-minute warning: Boffins train ImageNet-based AI classifier in just 240s

Faster is always better in AI, although it comes at a price. As researchers strive to train their neural networks at breakneck speeds, the accuracy of their software falls. Thus, teaching machine-learning code at high speed, with the resulting network able to make precise and correct decisions, is a vital goal. To that end, a …

  1. Anonymous Coward
    Anonymous Coward

    When you're all out of smart...

    just go big.

  2. JeffyPoooh Silver badge
    Pint

    "...balance your software....the optimal batch size..."

    "Balance" and "optimal" are keywords that often imply a knowledge gap and/or omitted subsystems. They can actually be understood as an admission of failure, or (if we're being generous) an unfinished project. They're probably nothing to be particularly proud of.

    E.g. One doesn't "balance" a petrol engine's spark timing, "optimizing" the trade-off of low and high RPM. No, no. Instead one invents the Timing Advance subsystem and thus provides ideal spark timing over the entire RPM range.

    E.g. The difference between a carburetor (with all its mindless assumptions) versus modern fuel injection (with all its sensors and feedback systems). The former may be optimized, but the latter can always be ideal.

    E.g. One shouldn't be satisfied with the "optimum" resistor value to poorly control the current in an LED. Rather one should recognize the need for a more complex constant current driver circuit, design it, and thus achieve ideal LED drive current over a wide range of input voltages.

    It's a different mindset.

    In this example, one might invent an embedded hierarchy of parallel multiple systems (i.e. merely memory for the neural parameters) covering a variety of scales. The training would also be paralleled, so little extra time. Conceptually a scaling framework placed over the present system. [Concept Needs Work Alert.] Then the final system could be ideal over the possible range of scales.

    Balancing, trade-offs, optimizing - these are actually red-flags that may indicate designers that are failing to imagine a system with one more control loop.

    Welcome to my thought processes...

  3. amanfromMars 1 Silver badge

    Cyber Crown Jewels ....... Casting Bands of Broader Light for Greater Beta AI Sees

    Thus, teaching machine-learning code at high speed, with the resulting network able to make precise and correct decisions, is a vital goal.

    I'd Venture IT Virtually Heavenly, Katyanna. And Practically Almighty. That's certainly a Great AI Goal.

    But 'tis nothing new, the Search for Discovery of the Holy Grails with Heavenly Tasks to Perform and Present. IT has been Baffling Humankind with Fantastic Goals to Achieve since Forever.And dangling AIMaster Key that Unlocks Everything to Inspection for Prevention of Inattention to Serious Abuse and Perverse Misuse, is a Touch of Manifest Genius which to admire and comprehend is Most Definitively OtherWorldly.

    No wonder it is so highly praised and prized whenever so immensely valuable.

  4. Divyas

    This just make me wonder how far ahs technology gone.

    We are evolving, but more than us, the tech around us is.

    1. DJV Silver badge

      Are you talking about the article or the post above from amanfromMars1?

    2. amanfromMars 1 Silver badge

      A NEUKlearer HyperRadioProActive IT Tangent via Novel Base Root Code to XSSXXXX*

      We are evolving, but more than us, the tech around us is. .... Divyas

      We are indeed evolving, but more than us, is the tech around us revolting too :-) When too hot to handle, lay back and enjoy the ride if you aint a target of particular and peculiar interest.

      *An AIdDriver Par Excellence.

      1. amanfromMars 1 Silver badge

        Re: A NEUKlearer HyperRadioProActive IT Tangent via Novel Base Root Code to XSSXXXX*

        An El Reg Exclusive?

        Some InterNetworking of Things are Just Too Critical and Vital and Criminal to Keep Deeply Hidden for Secrets More Generally Unknown.

        To

        boris.johnson.mp@parliament.uk <boris.johnson.mp@parliament.uk>

        Sent: ...

        3 August 2018 4:25 PM

        Subject: ....

        RAW and Rare Augmented Virtual Realisation

        Hi, Boris,

        A Little Something Great British for Renegade Rogue Rebel Tories to be Aware is Readily Available from and for ..... well, Right Royal Loyal Knight Sources with Almighty Forces would be a Universal Asset and Heavenly Gift, methinks. Where does your force come from and what/who does it feed and seed/activate and mentor? ...... http://amanfrommars.blogspot.com/2018/08/180802.html

        Hostile Parliamentary Coup D'État Territory and Virtual Safe Space Haven. What's not to like for AI NEUKlearer HyperRadioProACTive IT Leading Program with Myriad Alternative Projects for Realising the Future with Presentations Today being Featured as Supply for Tomorrow.

        Global Peer Review of its Fully Transparent Directions will ensure Crashes and Conflicts are Easily Avoided if Aired/Mooted/Diverted.

        And please be not surprised if word of this approach is enquired of you. .... for fresh up to the minute news of your action and future intention towards it.

        Regards,

        ....... which automatically returned a totally unsatisfactory ....

        The Rt. Hon. Boris Johnson MP

        Member of Parliament for Uxbridge and South Ruislip

        Thank you for contacting Boris Johnson MP. Please note that this email account is for constituency and other enquiries in my role as Member of Parliament for Uxbridge and South Ruislip.

        If you are a resident of Uxbridge and South Ruislip, please ensure that you have included your full name and address. No emails will be answered without full address and contact details.

        UK Parliament Disclaimer: This e-mail is confidential to the intended recipient. If you have received it in error, please notify the sender and delete it from your system. Any unauthorised use, disclosure, or copying is not permitted. This e-mail has been checked for viruses, but no liability is accepted for any damage caused by any virus transmitted by this e-mail. This e-mail address is not secure, is not encrypted and should not be used for sensitive data.

        But no matter, 1st Class Special Delivery is hereby Assured and Guaranteed Faultless.

        "Oh what a tangled web we weave. When first we practise to deceive" .... Is Boris a Good Spider or Bad Actor? Over now to the AIMaster Stage for his Interpretation and Intervention and Presentation of Virtually Remote Developments. :-)

        Boris the Spider

  5. Chris G Silver badge

    Progress

    So these guys have managed to train it to be 'not very good' in less than 40% of the time that a previous attempt took?

  6. Duncan Macdonald Silver badge

    Very low performance

    To be useful an image classifier needs well over 90% accuracy. If it cannot manage 90%++ then the number of incorrect results will make the system unusable.

    1. JeffyPoooh Silver badge
      Pint

      Re: Very low performance

      DM suggested, "To be useful an image classifier needs well over 90[++] % accuracy."

      Depends on the application. And you need to considers both missed detection (false negatives) and false detections (false positives), considering the expected rarity of the event, using the methods of Bayesian probability analysis.

      An intruder alert system, where a detection event merely causes the human guard to put down their magazine long enough to glance at the video monitor... Such systems might be very useful even with a false positive rate of 50%. But they'd better not miss anything, so 99.9% detection rate.

      Scanning a stadium looking for known terrorists, where the police snipers start shooting at anyone with a beard, forget about it. The failures will be spectacular, and the system managers have become the terrorists.

      It's a bit more complicated than a single "90%" descriptor.

  7. MaldwynP

    I've invented an algorithm that will scan a 100000 photos in 2 seconds with a 99% accuracy. It runs on two networked Pi devices. Oh! You want me to prove it? Never thought you'd ask me that.

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

Biting the hand that feeds IT © 1998–2019