Scale up and learn from massive datasets

- If you've been in technology for a while, then you probably remember the rise of big data. The term came from a NASA report. The agency described a "big data" problem. The problem was that widespread use of the internet and smartphones created enormous amounts of new data. The worldwide web contained trillions of words, billions of images, and billions of hours of video and audio. Now the challenge was how can companies get value from that data? It was almost as if these companies landed on a continent that was filled with millions of acres of wheat, but they didn't have any of the tools to bring that wheat to harvest. Then, around 2011, organizations started investigating the use of deep learning. Deep learning was a machine learning technique that used an artificial neural network that was many layers deep. Imagine that you had two enormous cakes, one chocolate and one vanilla. You needed to cut these cakes into thousands of slices so that each person got a slice. Each one got to choose either a slice of chocolate or a slice of vanilla. One of the most efficient ways to deal with this big cake problem is to have 100 helpers. These helpers would organize themselves into layers. Each helper would make a binary decision: Is it chocolate or vanilla? Then they would pass a slice from one layer to the next and make sure everyone got the right cake. Anytime you have hundreds of cake helpers with thousands of slices, you're going to run into a few errors, so the network would need some way to make adjustments. Maybe someone got the wrong slice, or maybe someone changed their mind. Every time there's an error, the network would go back and look for ways to improve itself. It would train itself. That way, every time there's a new cake, it would get better at making sure everyone got the right slice. At first, these deep learning neural networks were used mostly to classify things. You could tune the network using tens of thousands of images of cats. Then, anytime you fed a new cat image into the network, it would be correctly classified. Later, in 2017, the same technology was used by Google to create AlphaGo. This is the system that Google created to play "Go", that was mentioned earlier. This system could train itself on enormous amounts of data, then it would find extremely complex patterns. Once it identified these patterns, it would come up with new strategies on how to win the game. Later, organizations started using these deep learning neural networks to do all forms of predictive machine learning. You could classify your customers based on how much they spend. You could create inventory control systems that would accurately predict when an item would sell out. Then you saw companies like Tesla and Google using this technology to deal with one of the largest big data problems of all. They used a deep learning system to create self-driving cars. These cars would analyze all the big data that comes from driving. Then they would look for patterns to make the best predictions on how to drive. What started out as an accurate way to identify images turned into one of the best ways to do any form of prediction. That's why, for over a decade, artificial neural networks have still been the dominant form of AI.