Sign in to continue learning
PyTorch: Fundamentals
24h15m
You've seen how PyTorch emerged from researchers' frustration with complicated tools, and why it's become one of the leading frameworks for building neural networks. But now, it's time to understand what you're actually building. Let's ground this with a problem, one that you're going to revisit later when you train a model to solve it. Say you work for a local delivery company that promises to get orders to customers within 30 minutes, and you may make those deliveries by bike. Last month, you were late three times. Your manager wasn't very happy. One more late delivery, and your job could be at risk. Now, a new order comes in, the customer wants something delivered seven miles away. Do you take the job? Can you get there in under 30 minutes? If you're already comfortable with the way neurons model linear equations, and learn by reducing error, you can feel free to skip this video. But if you're new to this material, and you'd like a quick refresher, stick around. So neural networks are great at solving problems like this, if they have the data. And you're going to tackle this one using the simplest kind of neural network possible, just one neuron. And since complex networks are really just many neurons doing the same basic operation, understanding how one neuron works gives you the foundation for all of them. And while this is inspired by biology, neurons in this context are just mathematical units. They're tools with adjustable parameters that shift to match the patterns in your data. It's more intuitive than it sounds. Let me show you. Here's some historical delivery data. Can you see an underlying pattern? 5 miles took 22.2 minutes. 6 miles took 25.6. Now imagine a customer wants something that's delivered 7 miles away. Let's plot the data, and let's see what's actually happening. Notice how the points follow a straight line. A good predictive model for this data would be, well, a line. If you know the equation for that line, you can then predict new values. And here's the key insight. A single neuron is just a linear equation with two parameters, the weight and the bias. And the math will look like this. That's just the equation for a straight line. So all the neuron needs to do is to find the right values for W and B to create the best fitting line through all of your data. And that search for those best values, that's the learning in machine learning. So the question then becomes, how does a neuron figure out which line is best? So let's just start with a rough guess. We'll start by guessing weight is 1 and bias is 10. So then if I try to predict the delivery time for 5 miles, I'll get 1 times 5 plus 10 is 15 minutes. But in the real data, we could see that 5 miles took 22.2 minutes. So I'm off by more than 7 minutes. Looks like I need a steeper slope. So let's try weight equals 3.4 and bias equals 5. 3.4 times 5 plus 5 is 22 minutes. And now we're almost spot on. We're just 0.2 minutes off. It's a much better fit. Just like I looked at the error a moment ago and I thought, hmm, the weight needs to go up and the bias needs to go up and I need to change my slope, the neural network makes those same kind of decisions, but using math instead of intuition. It starts with random values for the weight and bias. And in PyTorch, these are called your model's parameters. Then it will look at how far off each prediction is from the actual data. The further off the points are, the bigger the total error. The network will then use calculus to figure out which direction to adjust the weight and the bias. It's essentially asking, if I increase the weight just a little, does the error go up or does it go down? And then once it's figured all that out, it will take a small step in the right direction, measure that error, adjust slightly, and continually repeat. Now it could do this hundreds or maybe even thousands of times, but eventually it will find values close to the best possible ones. And in PyTorch, you'll see all of this happen with just a few lines of code. But then you might be wondering, what happens when you start connecting thousands or even millions of these neurons together? Does the math get impossibly complex? Well, here's the surprising thing. It stays linear. Each neuron is still doing that same basic calculation. Let me show you how a neuron handles multiple inputs. Our delivery problem neuron had only one input, distance. But what if you wanted to consider multiple factors, say, distance, time of day, and weather? A neuron with three inputs will just extend the same pattern. It's all still just a linear equation. It's just more terms being added together. Each input gets its own unique weight, they all add up, plus the single bias value. And that's really all a neural network is. Neurons connected to neurons. A layer is simply a group of neurons that all take the same inputs. When you connect one layer's outputs to the next layer's inputs, you've now built a network. The first layer can take in your raw data, like the distance, the time of day, and the weather. The last layer gives your prediction, like the delivery time. All those layers in between, we call them hidden layers, because you never directly set or see their values. Soon, you'll see how easy it is to stack these layers together in PyTorch. But for now, let's just continue with just one neuron. In the next video, you'll see this complete pipeline and how you'll go from raw data to a deployed model. It's the systematic process that every PyTorch project follows. After that, you'll build your first neural network, and you'll train it on this exact delivery problem. Just a few lines of PyTorch code will bring all of these concepts to life. So let's keep going!