Articles Blog

How Does Your Phone Know This Is A Dog?

How Does Your Phone Know This Is A Dog?


Nat: Hi, I’m Nat.
Lo: and I’m Lo. And this is our 20% project where we go find out about
different Google stuff we’re curious about. Nat: Last year,
we got to make a documentary about how voice search works. And that was the first time
we heard about something called
machine learning. Lo: Since then, we’ve been
kind of fascinated by it. Nat: So today we’re talking
with Greg and Chris to learn more about it. How would you describe
machine learning to someone? Chris: Well,
there’s a lot of problems that’s really easy to solve
with computers. Computers can go and, like,
simulate how galaxies move and how the courses
of asteroids going and how close
they’re gonna come to Earth. I could never
hope to go and do that, but I can go
and do this problem, like, recognizing that that’s a tree, which is so much tremendously
more difficult for a computer. Greg: It doesn’t feel like
intelligence to us, because it’s so effortless
as a human being to do it, but for a computer,
it’s actually really hard. Lo: Because the real world
is kind of messy and unpredictable at times, the strict logical rules that go into traditional programming
just don’t work. Chris: Instead of
going and writing programs that solve the problem,
we write programs that learn to solve the problem
from examples. Greg: And it’s
this process of learning that allows them
to improve over time and to actually be more clever
than they would be if we wrote down
a very rigid set of instructions for them to follow. Nat: Machine learning is in so many different things
that we use today that it’s kind of like
this invisible magic ingredient. Greg: Phones with the ability
to understand human speech. Nat: Machine translation,
email spam filters. Lo: When you go to the ATM
and you give a check and it can read the handwriting. Nat: Or a photos app that can automatically
organize your photos based on the things that
you like to take photos of. Like, “Here are all
my mountain photos. Here are all my food photos.” Lo: “Here are all my
dog photos.” Nat: “Here are all my
feet photos.” Chris: Everything
from facial recognition to going and trying to recognize
whether a particle’s present in a particular collision
at the LHC. Nat: Which–we would just like
to take a time-out to say– is that big tube in the ground
over in Europe that smashes particles together
at really fast speed so that scientists
can use that information to unlock
the mysteries of the universe, among other cool things. Now back to our episode, ’cause I am getting
really dizzy doing this. Whoo.
Whoo. Lo: So researchers
and scientists are still experimenting and trying to find the best ways to teach computers how to learn. But a lot of the progress
is coming from these algorithms that are based roughly
on how the human brain works. And these are called
artificial neural networks. Greg: So the artificial
neural network is something that it’s
a rough mathematical cartoon of how a biological
neural network works. In a biological brain, we have individual cells
called neurons. Each neuron looks at
what its neighbors has to say and then decide
what it wants to say. And in artificial
neural networks, we have little
mathematical functions. We put them
in some organized structure, and then we say, “Okay, you guys all together
learn to do this task.” Chris: We have
lots of neural networks that are really great at
going and recognizing, you know, this is a cat; this is a dog;
this is a frog; this is a mouse; this is a horse; this is a
truck, and things like that. Nat: So take for example trying
to recognize what this is. Chris: So it used to be
that we were really proud if we could get a neural net
with three layers to work. And it’s recently that
we’ve made a lot of progress on techniques that allow us to
train much deeper neural nets. Nat: And that’s why
this kind of machine learning is also called deep learning. A neuron in the bottom layer
is just gonna be looking at a tiny piece of the picture and making some computations
about it. It doesn’t understand anything
about dogs specifically. Greg: But what the neuron does
understand is, it says, “I’m giving a signal
that’s useful for somebody “who’s giving a signal
who’s giving a signal who’s giving a signal
who’s giving a signal.” Chris: They’re kind of
able to unfurl this really
high dimensional knot and pull it apart and make it easier to go and,
you know, separate different things
that are– are close together
on the surface from things that are– were tangled
all together earlier. Greg: But then at the top, we’ll put two neurons. And these neurons look
at the whole picture so far. Nat: They’re basically experts
at making the final call, figuring out, “Oh, “all the layers below me
said these things, “so I know that this is a dog
or at least I’m 92.4% sure that it’s a dog,
so it’s basically a dog.” Lo: And while there’s been
a whole lot of progress with teaching computers
to learn, they’re still much slower
learners than we are and they make mistakes
that you and I wouldn’t. Chris: So what I was working on is sort of trying
to find a way to go and look at a neural net
and ask, “What does the neural net think
the platonic ideal of a cat is or the platonic ideal of a dog
or anything it can classify”? And suddenly, you can go
and ask, you know, “Neural net, what do you think
cat looks like?” You get a picture of a cat or,
you know, “Neural net, what do you think
a barbell looks like?” These weights. And it goes
and it shows you a picture not only of a barbell, but
of an arm attached to a barbell. So the model thought
that the arm was, you know– It only learned
what barbells were from looking
at pictures of barbells, and they’re often held
by muscular weight lifters. And so–so it learned
that there was– You know, barbells
have arms attached to them. Greg: It takes them
a long time to learn. You show it
a picture of a school bus. [horn honks] When it’s early in learning,
the very next time you show it a picture
of a school bus… [horn honks] it’s only
a little bit more likely to say “school bus.” It doesn’t get it even though that was
the very last thing it saw. Whereas, you know,
you say to a kid, you know, “That’s a filing cabinet.” And then a second later,
you say, “What’s that?” He’s not gonna be like,
“Shoe,” right? You know– Lo: Well, like,
the more you describe this, the more I’m amazed
by human beings. Greg: Yeah. Lo: Like, the effect is,
like, “Wow. Human beings
are really amazing.” Greg: We’re really amazing
learning machines. [soft squeal] [low grunt] [mimics bark] [soft squeal]

100 thoughts on “How Does Your Phone Know This Is A Dog?”

  1. Parag, here’s the documentary that we worked on about how voice search works: https://www.youtube.com/watch?v=yxxRAHVtafI

  2. I had never thought of the computer "neurons" as code functions. It seemed to me as if they'd need to use some sort of transistor that can change size to simulate neurons. Cool!

  3. #NatAndLo as always Learnt so much more about "Machine Learning" & "Deep Learning"

    And we would also love to know about where your skill sets are applied at Google.
    (burning question: what do you do at the Big G?)

    Would love to know the story behind the major Google Logo re-branding, please make a video on that.

  4. How do I join the AMA event scheduled for 1 pm PDT on Sep 25th.
    The link above takes to the calendar event AND the calendar event does not have details.

  5. What do you think a full body, transport sport, that works strength-endurance – balance for a life time- right out the door looks like? Oh yes and is one of only two, of the highest human output race sports on earth. And born on an American Iconic device. Also born at the hands of one person. Also the only race sport that will ever have begun in the daredevil theater and have evolved into the endurance theater? Oh yes and of course the innovation remains the same as long as humans remain in the physiological form as they are. Solve for X? Is X is sedentary lifestyle then why do we indoctrinate the US public with football , baseball soccer , hockey and basketball? Would be like teaching typing and clerical skill sets instead of teaching code? No one will ever use field sport movement language after age 19 for any daily use. Transport sports. SpikeBoarding is one of the greats…. and it was born right next door to Google NYC.

  6. Awesome video guys! Thanks for posting. You really explained the process very intuitively! Truly inspirational, should be shown in schools!!

  7. Hehe…. Thanbk you for making me feel a little more clever…. We need more of those videos that explain VERY complicated knowledge.

  8. agradesco a Dios por este gran paso en el desarrollo de nuevas tegnologias ,) no cave dudas que los seres humano somos grande y que cuando queremos podemos lograr cosas tangrade como esas ,muchas vendiciones equipo google. Dios los ilumine siempre soy muy feliz potodo es to gracias…palante.

  9. who put those people to talk? they are all wrong.
    Right now computers can learn all languages and recognoice all type of pronunciations in few years, how much it would take to a human to do that?

    If you show to a baby a bed for the first time, and then you show them another bed, he will not recoignice what is that.. is only after hundred or thousands of times after see, hear and touch things that we learn about them.

    You can make a neural network with many visual, hearing and touch inputs with less definition inputs than our senses with its own movement devices, and the machine learning will learn faster than a human to recognize things.
    But they are very carefull to not said this to the public to not alarm them.
    Of course we still need to cross the conscience issue that is our drive, but we can give the machine goals now and it would not need a conscience to learn.

  10. The reason we are such amazing learning machines is because we were not designed by evolution but by an all encompassing and compassionate God who cares for us. 😉

  11. I'm a Computer Engineering student, and I did a course in AI, we studied neural networks and even had to design one to classify footprints… But this video was such a great explanation, would have saved me a week of puzzling through a textbook 😛 thanks for the great content !!

  12. This is beginning to frighten me not because of errors that the machine "unwittingly" makes but because of those which have deliberately included. It would be so easy, for example, when searching for taxation regimes to omit one of them, which might just be what a government is looking for, but which is "politically incorrect" and which this government is currently unaware about. Such biasing should be eliminated from the machine-learning program by the use of a security checker to search for any introduced "pre-programmed" avoidance of some information of specific kinds of what otherwise appears to be a wonderful technique for sorting out data, otherwise it will soon have us all living in a "brave new world" or even "matrix" situation where we cannot escape from the system!

  13. Hey, that's Chris Olah right there ! I hope you appreciate him, he is a genius. So lucky to get to meet him just like that.

  14. was searching machine learning on youtube . found this video . Ended up watching every video in your channel and its around 4 am here now . guess my sleep is late, today too 😜.

    Awesome stuff guys . Keep going . 👍👍

  15. Hi Nat and Lo! I really like this video that you've created to explain this topic. More importantly, I think that it rocks that one of the top ranked YouTube videos on Deep Learning is by two gals 🙂

    I say that for a reason – my team and I just launched a series on Deep Learning ourselves here on YouTube. While we've gotten a lot of interest, the analytics tell us that the male to female ratio for viewers is 97% to 3%. Even considering the different things would you normally consider with stats, that is not a very good picture. Sadly, it makes sense given the anecdotal knowledge about this field – it is male dominated.

    We created this so it can help everyone. But if the last two weeks are any indication, letting it grow organically is not going to work. So, I wanted to pick your brains and see if you have ideas for us. How can we engage female audiences with our content? I would totally appreciate your thoughts on this :-)!

    Jag

  16. The A.R.B GO System + AlphaGo v Lee Sedol – Seoul 2016 – Prediction
    https://www.youtube.com/watch?v=MabOjRtU0kA

  17. BEST machine learning BOOKS EVER
    Python Machine Learning
    http://amzn.to/1UHC2X2
    Machine Learning: The Art and Science of Algorithms that Make Sense of Data
    http://amzn.to/1S9VsS7
    Fundamentals of Machine Learning for Predictive Data Analytics: Algorithms, Worked Examples, and Case Studies (MIT Press)
    http://amzn.to/1S9Vrxp
    Data Science from Scratch: First Principles with Python
    http://amzn.to/1UHC3dl
    Machine Learning: A Probabilistic Perspective (Adaptive Computation and Machine Learning series)
    http://amzn.to/1PiMoFZ

  18. Nat and Lo:

    As someone going to school in the fall for computer science at SDSMT, a massive fan of everything Google stands for, and a person wanting to work hard to become a part of the Google team I envy and look up to so much, I love these videos. I've already read 'Work Rules' by Laszlo Block and these videos give me even more reason to work even harder. Thank you for giving people like me a chance to see the great and amazing things you people do.

  19. I'm working on a similar process of using Psychology, Sociology, Philosophy, Humansim, Statistical Analysis and Philosophic Logic parameters, definition, relations, correlations and functions to create a new language which would literally be translatable between human and machine, all the way down to emotion, confounding variables and divine intervention as parameters included. It sounds like the two projects would merge well together. I understand the sciences of human behavior and the the mathmatics and logistic language symbolism behind how and why we function. I also have a slight background in computers. I've been looking for someone with the machine language skills to partner with me.

  20. I'd LOVE to share this video via our company blog. However, we are having troubling sharing it with the team in China. Would you mind if we copy it to a Chinese video platform?

  21. we (humans) are amazing learning machines. but it took us 2 million years to evolve to this. it's only taken computers 60 years to get to where they are. imagine in 20 more years or 10 or beyond.

  22. Good explanation. I like how they compare human learning with machine learning. With machine learning there's whole lot of human knowledge that goes into the design of the network and the supplying lots of examples. Human learning is orders of magnitude better than that. AI is great but don't look for the singularity in time soon because achieving that singularity requires so much more that merely expanding computing power.

  23. I know this comment isn't related with the topic of the video (the neural networks are amazing by the way) but I have to say that Greg (the man with a black t-shirt) is so so so hot for me 😂😂 pls I'm in love with him hahaha do you know his social networks… 🙂 congrats for your amazing channel

  24. Los que Des nos no res con na. Las pes ras. Por favor nos gas los vides nos con ras más. Dos. Nos que ros ves. Nas nes mi pes FES si res nes. Casa nos. Si gas no como si nos si nos Ges nes Des no si nes visi tas na dos la por les más. Gos nos más. Con Mos les pes nos res en para si Vi ras que. Gas ras ves en nes nos Des las Mos nas si con mos pos dos na res las ras. Des mes gas nes vos ma Des nos mes en Que los pes nos res las. Si por si gas pos Cos nos na uno mes Ges nes. Si que ras los nes nos dos por que por mi nos. Que kes si no nes dos nos. Los res las. Si mes los de gos si. Na ras. No no si les dos ños que con pasa ñas na para res los que los nos que los res Ges nes no que nes si

  25. OH MY GOD, what is that SCREACHING loud noise with the video? Seriously! Can't be just my computer. Any ideas?

  26. Just tried to explain the math behind neural network to my friend!! Imagine your fingers touch the table, they send signals to different parts of your body (coefficients), and different parts of your body have different pain thresholds so they respond to the signals differently (weights); when you push too hard and break your fingers, the pain lingers (feedback loops) ;))

  27. "Импортные черный фермы подняли икра." – that's fantastic on 1:34 )))) non-sense meaning in russian)

Leave a Reply

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