Slashdot Mirror

← Back to Stories (view on slashdot.org)

A Small Team of Student AI Coders Beats Google's Machine-Learning Code (technologyreview.com)

Posted by msmash on from the setting-new-records dept.

Students from Fast.ai, a small organization that runs free machine-learning courses online, just created an AI algorithm that outperforms code from Google's researchers, per an important benchmark. From a report: Fast.ai's success is important because it sometimes seems as if only those with huge resources can do advanced AI research. Fast.ai consists of part-time students keen to try their hand at machine learning -- and perhaps transition into a career in data science. It rents access to computers in Amazon's cloud. But Fast.ai's team built an algorithm that beats Google's code, as measured using a benchmark called DAWNBench, from researchers at Stanford. This benchmark uses a common image classification task to track the speed of a deep-learning algorithm per dollar of compute power. Google's researchers topped the previous rankings, in a category for training on several machines, using a custom-built collection its own chips designed specifically for machine learning. The Fast.ai team was able to produce something even faster, on roughly equivalent hardware.

29 comments

  1. Anyway, they were students by Tough+Love · · Score: 1

    They were students. Now they are Google employees.

    Good luck, you will need it.

    --
    When all you have is a hammer, every problem starts to look like a thumb.
  2. AI Algorithm? by 110010001000 · · Score: 3, Insightful

    How does an AI algorithm differ from a plain old algorithm. I am so curious...

    1. Re:AI Algorithm? by Anonymous Coward · · Score: 1

      Its Buzzword Factor is higher.

    2. Re:AI Algorithm? by Anonymous Coward · · Score: 0

      AI Algorithm (there is more than one) is SUB-TYPE of algorithm
      like sort algorithm (there is more than one)
      or search algorithm (there is more than one)
      or compression algorithm (yes you guessed it ... more than ...)

      for people working in computer science this is important info, "new algorithm optimization for AI allows 3 times faster performance" means much ore than "algorithm optimization for some unspecified algorithm allows 3 times faster performance", in first case you will remember "if i do anything AI related i should see if its applicable and use it in my work" in second case you will be "hmm there is thousands of algorithms used in my work, one of them is optimized each day and i don't personally use 99% of them anyway ... i will not waste time checking it out"

      car analogy: researchers optimized design of one of transport devices so it uses 50% less fuel by adding $100 device to its motor, it does sound really great, but i bet you would like to know if
      it was 50% less gasoline used for gasoline car (makes your car cost half as much)
      or 50% less kerosene used for airplane (makes your airplane ticket cost 50% less)
      or 50% less coal used by one of 3 coal powered trains still used in mid-west that will reduce your cost to send letters to Ohio by 0.1 cents

    3. Re:AI Algorithm? by Anonymous Coward · · Score: 0

      The same way a sorting algorithm differs from a search algorithm.

    4. Re:AI Algorithm? by Anonymous Coward · · Score: 0

      AI Algorithm (n). Any algorithm that the creator explain how it works.
      ---

      Seriously though, "AI" generally means it uses pattern matching against highly compressed seed / training data accumulated via a feedback loop.

    5. Re:AI Algorithm? by 110010001000 · · Score: 2

      Impressive. So it is like an algorithm, BUT also has Artificial Intelligence. I wonder what makes them think their algorithm is artificially intelligent.

    6. Re:AI Algorithm? by Anonymous Coward · · Score: 0

      How does an AI algorithm differ from a plain old algorithm. I am so curious...

      I'm not an expert in AI, but if you created an algorithm that did elaborate curve fitting F(x)=a+bx+cx^2+dx^3.... using countless runs then that is well likely called curve fitting, but if your situation is complex enough it is called AI I suppose. Somewhere in the mess the computer needs to be told what is a good outcome, or a better outcome, then training and iterative approaches allow it to eventually "learn" a potentially good, though likely not optimal solution.

      I, personally, would like to see more studies in finding optimal deterministic solutions to problems. The more problems we can solve in ways that are provably better, the more we can use that stack of solutions to better solve large problems. Basically you can do more if you broaden your tool base, rather than relying on magical solutions. Determinism first. AI and similar last.

      Certain things need solutions that are as deterministic as possible, such as, when operating a vehicle there should be, at minimum, some supervisor program that does nothing more than check that we are operating safely. If you could ever deterministically prove such a supervisor exists, then bam, you've solved the hard problem, and anything the AI does, while possibly stupid, won't kill anyone.

    7. Re:AI Algorithm? by Aighearach · · Score: 1

      How does an AI algorithm differ from a plain old algorithm. I am so curious...

      The use case.

    8. Re: AI Algorithm? by Anonymous Coward · · Score: 0

      Let's say ML = AI

      Factorial is not an ML
      algorithm.

      Newton-raphson and backprop are algorithms used to build ML algorithms.

      Regression and perceptrons are ML algorithms.

      The reason is because it is possible that every time they see an input, that input affects changes the return value of that input if it were seen again. That's to say that these algorithms maintain internal state based on observations of external state aka "learning".

      Systems that can learn complex tasks in anything resembling a general case tend to be called AI, and I personally feel this is because general cases requires multiple algorithms, each providing a "skill" to be combined into a more abstract problem solving ability.

      For example natural language generation requires a bunch of low level skills to be assembled, and the group of algorithms used tend to be more complex and multifaceted.

    9. Re: AI Algorithm? by Anonymous Coward · · Score: 0

      As the algorithm depends on training data, a unique distinction. What sets your own brain algorithm apart from this AI algorithm? Beyond a lot of hand waving, perhaps very little. The more we learn about brain representations, the requirement for magic sauce is growing ever smaller.

    10. Re: AI Algorithm? by Anonymous Coward · · Score: 0

      No need to make it either magical or very complex sounding.
      The talk about "AI" these days, is mostly about new inventions in ML and specialized hardware for running mathematical formulas (ie. GPU-farms and TPUs).
      Classical programming is about the programmer understanding requirements as input, and then crafting a program that takes input and produces output after some processing in the code. The most natural way for crafting programs is to start with a MVP and iterate to improve one's skills and the system itself, in order to best satisfy the environment over time.
      ML is mostly about an algorithm being fed input and desired output, then iterating over processing-output, gradually improving output to match desired output. It is said to "work" when the algorithm can be generalized to non-training data, which is a criteria for usability. ML is appropriate when it would be too hard/complex for humans to write the algorithm, like for image/letter recognition, so we make approximations using lots of computing, that sometimes may break catastrophically instead. You can look forward for nuclear arsenal to be controlled by ML in the future. God Save Us All!

    11. Re:AI Algorithm? by Anonymous Coward · · Score: 0

      Which one uses a hat?

    12. Re:AI Algorithm? by q_e_t · · Score: 1

      There are many sorts of neural networks. Using FP for weights isn't necessarily required, although it's certainly simpler to program backprop that way than try to use integer mathematics. Backprop isn't the only form of neural network available, although to some extent it depends on how you define neural network.

    13. Re:AI Algorithm? by dinfinity · · Score: 1

      Technically the students didn't really write or design an algorithm that works well in the benchmark. That very fact is what sets 'AI' apart from algorithms. You could argue that the students designed an algorithm that produces an algorithm that works very well in this case, but it is then still the resulting algorithm that works well for the task at hand, not the system they designed that 'came up' with the algorithm.

      I'd say that calling these things 'emergent algorithms' would be more appropriate or even better: 'machine learning systems'

      Let's see how that looks: "Students from Fast.ai, a small organization that runs free machine-learning courses online, just created a machine learning system that outperforms code from Google's researchers, per an important benchmark."

      Much better.

  3. Re:Sounds Gay by Anonymous Coward · · Score: 0

    TDS is real, seek help, for everyone else's sake.

  4. Re:Sounds Gay by Anonymous Coward · · Score: 0

    Mueller feels fine, but if you'd like to kiss Putin's dick I think he's still accepting that from Trumptards, you treasonous faggot bitch. Hang.

  5. Apple 1 by AlanBDee · · Score: 1

    Fast.ai's success is important because it sometimes seems as if only those with huge resources can do advanced AI research.

    Any one who believe's that needs to be hit over the head with a replica of the Apple 1.

  6. Heaven forbid they form a startup. Google buyout. by Seven+Spirals · · Score: 2

    Well, if you can't beat them, buy them. Kids, whatever you do don't go to work for Google. Make them buy you out, don't give them your skills for salary. Squeeze these corporate fuckers as hard as you can for the edification of the rest of us who didn't write a great algorithm that beats the Googleplex-of-assholes.

  7. No incredible. by Anonymous Coward · · Score: 0

    Then, why don't they publish their fast algorithms?

  8. Re:Punish my gaping asshole by Alalalalalalalalalal · · Score: 0

    Al Coders would readily sodomize you once Al Qaeda is done beheading you for your wicked sins, foolhardy infidel.

  9. Google has an incentive to be slow... by Anonymous Coward · · Score: 0

    They sell cloud GPUs by the hour, the slower Tensorflow is, the better for Google. Use PyTorch instead.

  10. Normal by Anonymous Coward · · Score: 0

    Few things come to mind:
    - It has been estimated by simulations and big research is best done if research money is spend evenly to every researchers, instead of focusing it to the best.
    - As Google had already the record, they had very little motive to try to beat it so they probably spend their time trying to do something else.
    - Google does grass root research, as in invent completely new things, their aim is to invent general AI, students on the other hand simply replicated Googles search and focused on optimizing it. So they have completely different goals.
    - Young people don't believe in limits. When I was a student, a teacher of mine showed me an example of code that is fully optimized and can not be improved any more. I took the challenge and with my friend managed to improve it more, teacher accepted our solution.

    It is really cool what they did, but not unexpected and it d oesn't prove that they are better than people at Google. It requires much less intelligence, wisdom and creativity to improve something than creating the original version. E.g. there were very few who invented air-plane, but once it was done, there were many who improved it.

  11. Re: Sounds Gay by Anonymous Coward · · Score: 0

    Shut up you faggot shill INCEL deplorable uneducated cis-hetero gaylord trumptard Russian NAZI alt-right bolshevik anti-Semitic Zionist cock-gobbling fascist mansplaining SJW shitfucker MRA trailer trash inbred lesbian Hillaryist feminazi ghetto alt-left white supremacist PEDOPHILE wetback spic mick wop nlgger chink kike redneck puritanical crackhead liberturdian commie TRAITOR!

  12. Re: Sounds Gay by Anonymous Coward · · Score: 0

    I envision fields of gold, surrounded by wonder and excitement, dripping with lubricant life force, filled to the brim with extraordinary excitement and soulful restfulness.

    Also, not many people know this, but cows are known for biting their own legs when they sleep.