back to article Python lovers, here's a library that will help you master AI as a newbie

If you’ve been thinking about trying to learn deep learning, here’s a new software library that promises to make things easy. Fast.ai, a startup co-founded by Rachel Thomas and Jeremy Howard, a professor and research scientist both working at the University of San Francisco, have released a free open source framework that …

  1. bombastic bob Silver badge
    Happy

    this has some potential

    An open source AI lib (etc.) for Python. This DEFINITELY has some potential!

    Additionally I'd like to see some API stuff for quantum computing included with it (if it hasn't been done already). All open source, of course!

    /me sees possibility of AI stuff running on RPi

    1. cbars Bronze badge

      Re: this has some potential

      an API for quantum computing? Need the hardware first!

    2. Sampler

      Re: this has some potential

      AI on an RPi? How long you willing to wait for that task to run?

      1. David Shaw

        Re: this has some potential

        I'd use the RPi as a front-end to all of the AWS instances that werent busy selling xmas/hallowe'en presents. The 'rules' would presumably be available after just 42 minutes?

      2. Tom 7

        Re: this has some potential

        There's already tensorflow on the Pi and on a Pi3B+ with a heatsink it can run the 'starter' Mnist train and verify in under 10min IIRC.

        I've seen a PiZero doing realtime object recognition on a pretrained model using the RaspberryPi camera.

        My only objection to this addition to the AI collection is saying its somehow easier to program in than say tensorflow. I'd bet all they are doing is just pre-setting some parameters that you will want to vary later so its not really any easier than tensorflow.

        And if you really want to master AI Weka takes a shit load of beating given you dont have to write a single line of code as the GUI can be used to set just about anything.

    3. Anonymous Coward
      Anonymous Coward

      Re: this has some potential

      https://github.com/Qiskit/qiskit-terra

  2. Lusty

    Maths

    "Many people think they need to spend years studying advanced math first [to learn AI]"

    Not true. ML and AI are already commoditised to the point that a toddler could train a model. Many people do, however, think that you need to understand maths to know if your model is doing what it's supposed to. If the "what it's supposed to" is any more detailed than producing an output, any output, then you'll probably want a keen understanding of maths.

    1. Anonymous Coward
      Anonymous Coward

      Re: Maths

      I think the days when almost all programmers were doing more than turning the handles on systems they do not understand to produce answers they do not understand are long gone.

    2. Anonymous Coward
      Anonymous Coward

      Re: Maths

      The rubber hits the road when someone has to stand up in court and explain why their super-duper AI doesn’t want to give mortgages or jobs or whatever to black people.

      1. Paul Hovnanian Silver badge

        Re: Maths

        "why their super-duper AI doesn’t want to give mortgages or jobs or whatever to black people"

        Or runs over them.

      2. Anonymous Coward
        Anonymous Coward

        Re: Maths

        Woah fella. You've got it all wrong, if an AI starts rejecting black people for mortgages and jobs it will have been licensed from the courts.

      3. bombastic bob Silver badge
        Childcatcher

        Re: Maths

        "explain why their super-duper AI doesn’t want to give mortgages or jobs or whatever to" [insert special interest group here]

        well, if race/sex/whatever wasn't an input parameter to your algorithm, the answer to this should be obvious. The moment your algorithm includes something that *could* be a matter of discriminatory practice, an improperly trained AI *could* make use of it in a discriminatory manner.

        Otherwise, FUD.

        Besides, wouldn't having a "simplistic" library, something that n00bs and kids can easily make use of, help people to understand the basics of AI and thereby avoid the FUD-trap?

        [I can't believe how my original positive non-snarky post about this got 8 downvotes - probably ALL from my 'fan club' - which means its personal and unrelated to content. howler monkeys. go fig. the up/downs are meaningless - well i wear the 'downs' as a badge of honor, heh!]

        Icon, because "think of the children" which is why I also mentioned RPi - when kids get this kind of AI API working with an RPi (in Python) along with whatever science project they're interested in, things could get really interesting.

        1. Anonymous Coward
          Anonymous Coward

          Re: Maths

          I guess you don’t really know very much about this field because a) you think ANOTHER free Python AI library is news (we are drowning in them) and b) protected characteristics such as race can very easily be inferred from other inputs (age + postcode + one other for example) with a high degree of probability. That is what you need the maths to understand and compensate for.

          That’s why you’re being downvoted.

        2. Anonymous Coward
          Anonymous Coward

          Re: Maths

          Having a 'simplistic' library is a bit like having a 'simplistic' version of general relativity or quantum mechanics. In all three cases it turns out that you need a fair degree of mathematical sophistication to actually do anything useful with the thing, and a simplistic version just ends up with the monkey-turning-the-handle-on-the-system-they-can't-understand problem. Of course, if what you want to end up with is 'here's this magic box, feed it some data and it will tell you some things' rather than actually understanding what's happening then that's fine, I suppose.

          1. Anonymous Coward
            Anonymous Coward

            Re: Maths

            In all three cases it turns out that you need a fair degree of mathematical sophistication to actually do anything useful with the thing

            In the case of AI you'd need a f*cking miracle to actually do anything useful with the thing. I'd imagine libraries like this just encourage the unknowing to come up with "a model" that "does things" and likely overfits the data or produces some other poor result someone well versed in the arts would not do.

            AI for beginners is what happens when everyone wants to jump on the AI bandwagon and there aren't enough specialists to go around.

  3. Mike 137 Silver badge

    Yet another opportunity

    Yet another opportunity to create unaccountable systems with potentially far-reaching and possibly adverse effects on the public at large without having to understand the first thing about what you're doing...

    1. SecretSonOfHG

      Re: Yet another opportunity

      "unaccountable systems with potentially far-reaching and possibly adverse effects on the public at large without having to understand the first thing about what you're doing"

      Sounds like how 99% of software is created, managed and mantained

      1. Anonymous Coward
        Anonymous Coward

        Re: Yet another opportunity

        "Sounds like how 99% of software is created, managed and mantained"

        Spot on. We had a customer SLA that was derived from a major product's network management overview of SNMP events. On closer inspection we found that in some conditions the underlying events were misinterpreted as longer breaks than had actually happened. We wrote our own processing which made sure that the knotty ones were highlighted for a human mind to untangle.

        In the past IT people often didn't understand the underlying software/hardware of their environment. Nowadays that complexity has increased considerably - and people's training/knowledge has become even more limited or specific.

    2. Dr Dan Holdsworth
      Boffin

      Re: Yet another opportunity

      My PhD supervisor used to derisively refer to what he termed "Statistical Stamp-collecting". This was 20 years ago, and even then it was possible to start off with a decent-sized biological database of a few thousand data points per treatment, and run an ANOVA analysis comparing each sample to every other sample, and do so whilst one nipped down the pub for an only-slightly extended lunch break.

      Actually working out what the results actually meant, that was the tricky bit, as was deciding whether or not the experiments were well enough designed to support the inferences you could statistically "prove".

      These days, of course, we have the famous meta-analysis. Not got the time or budget to do work yourself? Need a few more papers published to go for that professorship? Easy, munge together several other groups' work without a thought to the rigor of each experimental method, fire it all at something statistical (Like Kruskal-Wallis so not bright spark can point out that the data aren't parametric) and hey ho, a-correlating we shall go.

    3. Paul Hovnanian Silver badge

      Re: Yet another opportunity

      So sad!

    4. FuzzyWuzzys
      Facepalm

      Re: Yet another opportunity

      Must be draughty up in the ivory tower Dr Sheldon Cooper!!

      God forbid some secondary school kids in a computer lesson get to learn some hands on basics about AI using something easy to get started like Python and a downloadble package for free.

  4. SVV

    All it takes is a few lines of code

    Well, if that's all it takes to "master" the art of creating true artificial sentient entities to do my bidding, my plan to create an army of cyber brains and take over the world will be operational by lunchtime.

    1. Lusty
      Terminator

      Re: All it takes is a few lines of code

      Unfortunately you also need to be a billionaire to affort the hardware on which to run it by lunchtime. My sentient cyber brain is at the angry toddler stage after a year of training on an old Pentium 4...mwahahahaha

      1. Anonymous Coward
        Anonymous Coward

        Re: All it takes is a few lines of code

        Yeah well, I've miniaturised my tech to fit in a Planters tin. It's just nuts.

    2. Anonymous Coward
      Anonymous Coward

      Re: All it takes is a few lines of code

      Could you wait until mid-afternoon? It's curry day in the canteen on Thursdays. :-)

      1. Anonymous Coward
        Anonymous Coward

        Re: All it takes is a few lines of code

        AI waits for no man. And doesn't need curry. Exterminate!

        1. hplasm
          Happy

          Re: All it takes is a few lines of code

          "AI waits for no man. And doesn't need curry. Exterminate!"

          exterminate! And put it in the curry!

          /Q8

        2. Anonymous Coward
          Anonymous Coward

          Re: All it takes is a few lines of code

          > AI waits for no man. And doesn't need curry. Exterminate!

          I read that as "AI waits for no nan. And doesn't need curry. Exterminate!" - still got curry on the brain!

  5. Anonymous Coward
    Anonymous Coward

    A few lines of code.....and.....

    ....what else?

    *

    Where does the "big data" database come from?

    *

    Oh....and there's the old conundrum that correlation does not say ANYTHING AT ALL about cause!

    *

    Just an old f**t here worrying about click bait on El Reg.

  6. disgruntled yank

    Master

    Master, or employ?

  7. Anonymous Coward
    Anonymous Coward

    Question re:mixed precision traininf

    “Behind the scenes, we’re following all of Nvidia’s recommendations for mixed precision training. No other library that we know of provides such an easy way to leverage Nvidia’s latest technology, which gives two to three times better performance compared to previous approaches”

    Is the mixed precision functionality available on consumer GPUs or are they locked to the professional cards?

    1. Clem H. Fandango

      Re: Question re:mixed precision traininf

      I believe Volta and Turing can do half precision for the speed gain, though older GPUs can still take advantage of the memory reduction.

    2. bombastic bob Silver badge
      Meh

      Re: Question re:mixed precision traininf

      hopefully the AI API has both GPU and non-GPU versions of the core functionality with a simple way to pick the right one based on your hardware. Otherwise why use Python? The kind of thinking that drives users to a specific hardware platform is WORSE than something that uses ".Not" or C-Pound (or requires you to log in as 'administrator').

  8. Anonymous Coward
    Anonymous Coward

    “gotten”

    Do Americans really use “gotten” when speaking/writing in a formal tone?

    I can never work out whether it’s a charming hangover from mediaeval English (that standard English has dropped), or deliberately cranked up annoying tweeness remixed by marketing weasels in order to sound cute and folksy...?

  9. wayne 8

    So easy a caveman can do it.

    "master AI - All it takes is a few lines of code"

    Let me guess, the library is an AI itself.

    Relevant, today's SMBC link:

    http://www.smbc-comics.com/comic/ai-3

  10. Paul Smith

    Bollock naked emperors

    "Many people think they need to spend years studying advanced math first [to learn AI]"

    As someone who has spent those years and learnt advanced data science (including the niche that is AI/ML), I initially laughed at the idiots who thought they could use AI/ML for all sorts of applications. Now I think that those years were wasted when I see PFYs being paid more then me to produce 'applications' with embedded AI's. When challenged to explain how they work (if they work) they claim it is too complicated for anyone but an expert to understand, when in fact it is usually just a pretty crude decision tree.

    1. Anonymous Coward
      Anonymous Coward

      Re: Bollock naked emperors

      If you think mastery and training lead to reward* you're very much mistaken and have probably been lied to at school. Those PFYs are still smarter than sales people and likely earning 1/4 of the money of sales people who have no qualifications other than a nice personality. A nice personality is easy enough to get, just spend your education years in the pub socialising rather than in a lab training...

      *Financial reward. As any underpaid teacher will tell you knowledge and fulfilment are their own rewards. Which is why they are underpaid.

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