Accurately predict what someone will do

- As humans, we try to learn from our own mistakes. From poor backgammon moves to bad barbecue. It's our mistakes that help us grow. In fact, many people believe that it's our ability to learn from mistakes that defines intelligence. So far, you've seen that artificial neural networks scale machine learning to work on massive data sets. Remember, that was part of our big cake problem. We used hundreds of workers or nodes to break down the cake into thousands of slices. But this cake neural network has a big challenge. It can process a huge amount of data, but it needs to learn from its mistakes. So you might have someone at the very edge of the network who's demanding a donut. Workers in this cake network might ignore this feedback and keep serving cake. That means to improve the system, feedback must go from the edge of the network all the way back to the beginning. The system has to tune itself to respond. In neural networks, this self tuning to improve the performance is called the back propagation of errors. When the network makes a bad guess, this failure goes back up through the network to make future changes. You'll often hear this called back propagation or back prop for short. A key part of a neural network is that it's adjustable. The nodes on the network need a system to adjust so that they can be trained. They need to be able to continue to improve and grow. So think of it this way. Imagine that you're an executive for an airline company. One of your biggest challenges is that you have high fuel costs. That means you want to make sure that you can fill your planes with the optimal amount of fuel. That way, the passenger to fuel ratio will make it easier for you to earn a profit. The more crowded the plane, the easier it is to fly people at a lower cost. But predicting who's going to be on the plane is a real challenge. You must sell a certain number of tickets, but some connecting flights will be delayed, or passengers might board an earlier flight. In this case, you could use a neural network to go through all the big data of your previous flights. It can identify patterns and then predict who will be on the next plane. You can then test the network based on some new data to see if it does a good job making a guess. Back propagation algorithms can then improve the accuracy of the artificial neural network by making small changes. If it's way off, then the algorithm will make a big change. Then the network will try and make a better guess. It can also make much smaller tweaks. With every new bite of data, this network is getting better at making predictions. It's getting feedback, and then improving itself over time. If you're a frequent flyer, you may have even noticed these improved deficiencies. A few decades ago, you may have seen a couple of empty seats. Today, you probably fly on many more full flights. That's because airlines can now use improved neural networks to predict the weather, travel habits, and seasonal fuel consumption. It's all about recognizing patterns in large data sets. Then neural networks adjust themselves to make more accurate predictions.