Tech Explained: Artificial Intelligence

Latest in our Tech Explained series, we find out what is artificial intelligence and how it is changing our lives.

Artificial Intelligence (AI) has been touted as a way to improve almost every aspect of life. There are already AI systems that can compose songs, monitor construction projects and detect heart attacks. What exactly is AI, and will it turn the world into a Terminator-style dystopia, or create an Eden where computers do all the work?

Put simply, AI is a term used to describe computer software that can learn from experience. This is very different from most software, which can only perform tasks it has been pre-programmed to execute. One way to see how AI is different is to think about the way humans recognise a dog. There are hundreds of different breeds of dog in the world, of all different colours, sizes and shapes. Yet humans, even very young children, can instantly recognise any of them as a dog. To teach a computer to do this it is necessary to first show it millions of photos of dogs. It must then be taught how to compare any new image it receives to the photos, and draw a conclusion as to whether the new image is also a dog. Achieving this requires two things – a lot of very large data sets and a lot of processing power to analyse the data. It is only recently that these two things have become readily available.

One of the most powerful types of AI software is called deep neural learning. In deep learning, many different algorithms (small software programmes) are connected together in layers. Each algorithm looks at a different part of the problem, and also weights the calculations made by the earlier layers. For example, in the problem of how to recognise a dog, the first layer might focus just on the outline, another on colour, and yet another on size. Every layer weights the input from the previous layers and then moves all the information on to the next layer. The result (dog or no dog) is based on calculations from the collective weights of all the layers. In this way, the software has taught itself how to recognise a dog.

For now, most AI can only preform a clearly defined task. For example, an AI algorithm that has learned how to play poker cannot also learn to play chess. The next step is to develop an AI system that can solve new problems or teach itself new tasks by applying information it already knows. This type of AI is called artificial general intelligence (AGI). Some progress towards AGI has already been made. Researchers at DeepMind have created an algorithm that has learned how to play both Go and Chess. The final step would be an AI that is actually conscious, that is self-aware. Will AI one day be able to do everything humans can do?