How Do Machines Learn?

How Do Machines Learn?

How do all the algorithms around us learn to do their jobs?
Discuss this video:



Thank you to my supporters on Patreon:

James Bissonette, James Gill, Cas EliΓ«ns, Jeremy Banks, Thomas J Miller Jr MD, Jaclyn Cauley, David F Watson, Jay Edwards, Tianyu Ge, Michael Cao, Caron Hideg, Andrea Di Biagio, Andrey Chursin, Christopher Anthony, Richard Comish, Stephen W. Carson, JoJo Chehebar, Mark Govea, John Buchan, Donal Botkin, Bob Kunz

How neural networks really work with the real linear algebra:

Music by:

You may also like...

69 Responses

  1. salim karimov says:

    what if there where three bees

  2. Irpuxa says:

    Awesome video. Very engaging! πŸ˜‰

  3. Sean Bouffard says:

    The algorithm that brought me here is named Philip Defranco

  4. Corn Bread says:

    Can I please get a plushie of the “special” student bots?

  5. Oliver Harman says:

    Im just commenting just because you told me to

  6. Modest Zlatolobov says:

    Bots made me to write this comment. |-|e1p

  7. EpicKaiserTom says:

    I’m just gonna plug 3blue1brown’s channel ( as he has an extremely in-depth series on how neural networks function down to the actual maths, and he works to try and ease you into it as best as possible.

    If you are interested in the nitty gritty I implore you to start there and get really interested in it before fully diving into the absurdly dense text and courses about it.

  8. The Twig of Newberry says:

    Nice ending

  9. Vapez says:


  10. Hugotistical says:

    “Teacher bot can’t teach, But it can test” Did you just explain the entire American School system?

    • Anthony Eaton says:

      The whole theme of “not actually knowing the core issue, merely succeeding on tests” really drives it home.

    • Tanner Shackelford says:


    • DeRockProject & the Attack of the Really Long Channel Name says:

      and if they fail the test, they go into the fire. sounds about right!

    • P E A C H says:

      Just 95% of it

    • Damir Liashchuk says:

      ❀❀❀ *Iсh bin sΘ― Δ—insam und suche dΠ΅n MΠ°nn dΠ΅s Π’rΠ°umΠ΅s*
      ❀❀❀ *Sc̣hrСib mir*
      ❀❀❀❀❀❀ 18+ ❀❀❀❀❀❀

      ❀❀❀ *Iс̣h bin so Π΅insΠ°m und sucΜ£hΠ΅ dΔ—n Маnn dΠ΅s Π’rΘ§umΔ—s*
      ❀❀❀ *SchrΔ—ib mir*
      ❀❀❀❀❀❀ 18+ ❀❀❀❀❀❀

      Phil N Good Shurwayne Stranger
      Fresh Epcot With Eric Darden Horticulture Manager Walt Disney World
      Kamikaz Seul ou en Equipe 2016
      L Aveu Fabien Cavalerie Justine Le Pottier

  11. Giovane Dias says:

    Sooo…. Maybe that’s why I need to keep re-liking the videos I already watched on my phone?

  12. Jackson Reno Kripperino says:

    gay (not actually gay)

  13. harisbiggy says:

    So basically the same as evolution.

    Teacher bot: Environmental factors (which through elimination, choose which organisms succeed).
    Builder bot: The random mutations that occur in every reproductive cycle producing unique organisms.
    Student bot: The organisms, all with a different, unique DNA code

    Only the best adapted organisms/code survive. In CGP Grey’s example it was testing the ability to distinguish bees and threes. In evolution it is testing the ability to pass on your DNA. Only the best organisms for their environment are allowed to continue on to the next cycle, where again the same tests happen, producing creatures that are better and constantly more adapted to their environment. As always, nature seems to have the solutions figured out and we’re just getting better at copying it πŸ™‚

    • Yash Raghuwanshi says:

      You just blew my mind!!!

    • columbus8myhw says:

      In fact, this specific type of machine learning is called a “genetic algorithm” or “evolutionary algorithm”. Other types of machine learning, such as neural networks/deep learning, work differently to this.

      EDIT: I should clarify, neural networks/deep learning can still be done with genetic algorithms. They usually aren’t, though; they’re usually done with a form of gradient descent.

    • Enourmousletters says:

      One important difference is we can alter what Teacher bot is testing for to get more nuanced results.

      That would be like if we could decide to extend an organisms lifespan not because of any benefit to the species as a whole but just because we felt like it.

    • DeRockProject & the Attack of the Really Long Channel Name says:

      The algorithm bots dont know what they’re doing. Their creators dont’t know what they’re doing. It’s the process of evolution, how our brains are made. How we’re made. Goddamn it…

    • Sergey says:

      That’s exactly what we do when we perform selective breeding on animals or plants. The most important part is not lifespan, but amount of influence it has on next generation.

  14. The PC Security Channel [TPSC] says:

    Isn’t that how evolution works too? Only slower.

  15. Isabelle Wolvetang says:

    This is kind of scary

  16. Crystal says:

    It’s so crazy that we’re building things made to be smarter than we ourselves are. It’s exciting but…also a little unsettling.

    • PogieJoe says:

      I’m even more unsettled by the idea that the machines don’t even understand how, why, or to what degree they are intelligent.

    • NabsterHax says:

      And then you realise the exact same thing applies to human brains.

    • Blazing Nipples says:

      Videos like these show us that we’re further off from those AIs than the current popular mindset believes. When that time does come, they will demonstrate their intelligence as a virtue, and we will voluntarily follow them. Because really, there is no us and them, as crazy as that may seem. The bots ARE us, and we are the bots. An intelligent being will see that, and many other things. There’s nothing to worry about. The real fear is limited AI. AI that is only good at one thing, but is super good at that thing. Never self aware, but superior to any human in a given task. The hope is that well rounded, smart AIs will be able to reign these limited AIs in. China is developing AIs similar to what youtube has, but solely for the purposes of political repression and censorship. This is a disturbing possibility for limited AI. On the other hand, limited AI can be used for specific medical tasks. So really, like all tools, its what we make of them. I think that AIs are our only hope, because humans are gonna nuke themselves if they keep running themselves for much longer. Our technology outpaced our ape brains somewhere around industrialization and the species can’t be trusted to hold its own guns anymore. Muskets and canon balls were about as much as we could handle. Anything more is giving a chimp a loaded handgun, and with the nuclear age, that guns safety was taken off. I welcome our AI overlords. The age of the meatbags must come to an end.

    • NabsterHax says:

      @Blazing Nipples. Right, I don’t fear the self-aware general AI of the future that can solve problems that we don’t even know need solving. It’s the idiot machine that’s only goal is something benign until it begins working out how to carve up humanity and the planet to better achieve its goal. It’s not smart enough to reason with, and hasn’t figured out that if it destroys the future for higher gains now it’s going to end up worse off.

      Wait a minute, that sounds familiar. Oh god. What have we done?

      AIs that are better than humans at improving humanity is humanity’s only hope.

  17. Jack Jackson says:

    Robotic Artificial Selection. Darwin applied to computation.

  18. Flamelight 7 says:

    This is terrifying and adorable and I love how dejected Grey sounds when he sells out to the algorithm bots.

  19. Jay Infinity says:

    And this is what Elon Musk is a afraid of. #BuildABot

  20. kanish161 says:

    this looks an awful lot like evolution

    • Trap Sensation says:

      There’s a field in Computer Science called Evolutionary computation, and it’s being used for machine learning. My friend who’s a research assistant at MIT believes that Evolutionary computation will be critical in the future for machine learning, so you’re definitely on to something.

Leave a Reply

Your email address will not be published. Required fields are marked *