When AI became Silent

When AI became Silent

When AI became Silent

For a while, it looked as though the problem of machine intelligence was finally being solved. Neural networks could learn. Algorithms could adjust their weights. Machines could recognize patterns from examples. But then something unexpected happened. The excitement began to fade. The computers simply weren't powerful enough. The data wasn't large enough. And many of the techniques that looked promising in controlled experiments struggled when the problems became larger and more complicated. Researchers had discovered something important: Knowing **how** to make a machine learn was not the same as having enough resources to make it learn something useful. Imagine trying to teach a child to recognize every object in the world using only a handful of photographs and a few minutes of practice. The problem isn't necessarily the child's ability to learn. The problem is the material available for learning. Neural networks faced a similar limitation. And there was another problem. Even when researchers knew an algorithm should work, the computers of the time could only perform a fraction of the calculations we take for granted today. Training a large neural network wasn't simply a matter of waiting a little longer. The required computation could be enormous. As expectations grew faster than the technology could deliver, enthusiasm turned into disappointment. Funding declined. Projects were abandoned. And artificial intelligence entered periods that became known as **AI winters**. The phrase sounds dramatic, but it describes something very real. When promises became larger than results, interest cooled. But the underlying ideas didn't disappear. Researchers continued working. And slowly, three things began changing. First came **data**. The world was becoming digital. Images, documents, websites, conversations and countless other forms of information were accumulating at an unprecedented scale. Then came **compute**. The graphics processors originally designed to render images and video turned out to be extraordinarily good at performing the kinds of mathematical operations neural networks needed. These became known as **GPUs — Graphics Processing Units**. And finally came better **algorithms and architectures**. Researchers were finding more effective ways to train deeper networks and extract useful representations from enormous datasets. Three forces that had once been weak were now becoming stronger at the same time. **More data. ** **More compute. ** **Better algorithms. ** And that combination changed the economics of learning. A neural network that would once have taken an impractical amount of computation could now be trained on vastly larger datasets. The machine could see more examples. Process more information. Adjust more parameters. And learn more complicated representations. The pieces were beginning to line up. But there was still a question. What kind of problem should we give these increasingly powerful networks? One answer came from something humans do effortlessly. We look at a picture... and recognize what we're seeing. A face. A car. A dog. A handwritten number. For a machine, that seemingly effortless act was extraordinarily difficult. Until researchers found a way to make neural networks particularly good at seeing patterns inside images. And in 2012, one system would make the world pay attention. Its name was **AlexNet**. And this time... the network was deep.