What better way to really solidify your knowledge of machine learning than with a fun game. This is that Quick, Draw, Q-U-I-C-K draw.withgoogle.com. And you're going to see how poorly Evan is going to draw some of these images and put you to see and hear the machine learning algorithm behind the scenes. Try to guess what it is that I'm drawing. I haven't seen these images before. So let's do this. All right, draw a donut in under 20 seconds. The first one's an easy one. I'll get through six. So let's do this thing. Donut, okay, based on what I know about a donut, yeah, there we go. All right, that was an easy one. Maybe we can get a harder one. Sword, that's even easier. I can totally do this. Maybe just like a stick, does it just get it from there? No, all right, let's add a little bit more down here. Okay, maybe we'll add a little bit of a thing up here. That's a perfect sword. No, guess it, come on, maybe someone needed to be holding it. Draw grapes, this should be easy, lots of circles and maybe like a thing on the top. Draw a dumbbell. Okay, okay, okay, dumbbell is something like this. How did you even get that from there? Okay, broccoli, okay, how do you draw broccoli and not draw a tree? No, no, I have no clue what you're drawing! [LAUGH] Erase it, erase it. Draw a dinner plate. That looks like an egg. No, all right, guys, I can't draw broccoli at all. All right, last one, draw matches. Now this should be okay. So we'll draw kind of like a candle. And then, How did you get matches yourself with that? Wow, it's impressive. How'd you not get that as a sword? That's incredible. Well, I hope you had fun drawing your six drawings and seeing me struggle through drawing with a mouse here in the studio. One of the things that I really wanted to highlight though is with any unstructured machine learning problem, which is like image data that you're feeding into a recognition system or structured data, which is where you're predicting values just on rows and columns of data, the key takeaway for all of machine learning is that you need to have good historical labelled training data. Labelled means you know what the right answer was in the past. And you can see that the best part of this whole website is that if you click on data, you can see the 50 million drawings that people have drew over time for, let's say like, what is this? That all the different ways people have drawn a basket, you can click on these. And the model has the benefit of continuously learning from each of these over time. So we can basically say, all right, I do know it's a basket. That's the right answer. And this is how I can arrive at that answer and I can make those predictions itself. Now let's take a look at another image. What's another hard one? A blackberry, yeah, that's a hard one. How do you draw a blackberry? You can see this, even humans have many different ways of drawing. How do you draw the difference between a blackberry and a grape? So even with just one color, you can see that this can be particularly difficult for machine learning model where it's like, is that a blackberry or is that a cherry? Everything else kind of looks like grapes as well too. Somebody started drawing it out like Bl for blackberry too. Man, yeah, you can see that, just a torrent of data so over 100,000 blackberry drawings there and you heard it trying to guess along live. Well, next up, we'll talk about the algorithm behind the scenes that looks at one of these pictures and then thinks about as a human does, all the different features of that image as it arrives at its ultimate conclusion. It's called a deep neural network. And we'll show you how you can think about the architecture of those and another tool called the TensorFlow Playground, where you can actually experiment with the deep neural network yourself. Good luck with the drawing and feel free to have more fun by trying to draw more and more images. Here's the algorithm behind the game that we just played. It's called a deep neural network or DNN. The layers are meant to mimic our own human brains in the way that we perceive stimuli. With each layer, a trained model learns more and more about the image of this dog that's hiding in a laundry basket, starting from things like basic detection of edges in the photo, and then colors, and ultimately arriving at a final decision of is this a cat or is this a dog? You'll get a chance to build your own image recognition model later on in the course. And spoiler alert, you don't need to write any code. But if you want additional intuition of how these model types like DNNs work, check out the TensorFlow neural network playground at playground.tensorflow.org. It was one of my favorite learning tools for when I was understanding how do computers actually think. And it shows just how far ML has come that we can build these models at scale into applications like Google Photos.
