A brief history of AI for non-techies

The "new" big idea, Artificial Intelligence, is not the brand new concept it might seem. I think showing how AI has developed makes it easier to understand how we might use it today. So, here’s my brief, high-level look at AI and how we got here.

The idea behind this software engineering tool has been in development for a century. During that time, engineering focus on it has ebbed and flowed. It’s had periods of intense activity and three or four AI winters.

It first grew from investigations into how brains work at the end of the nineteenth century. This early diagram shows how they recognised the brain as a network, not just a lump of grey matter. In the 1940s McCulloch & Pits published a paper on brain cells and the wiring that includes neurons. This idea progressed in biology, but also found its way into engineering. By the 1950s, the increase in computer power meant that organisations could simulate hypothetical neural networks, resulting in a commercial product: MADALINE which eliminates echoes on a phone line and is still in use today.

But in the middle of the twentieth century AI entered a winter. Progress slowed because the technology couldn’t keep up with the early promise, we couldn’t replicate the Exclusive OR Gate; a fundamental building block in electronics and they were just not giving the results we expected! The idea was reignited in the 1980s when a new training technique called backpropagation allowed us to create "hidden" layers of neurons. Backpropagation is arguably the last new science in AI. It allowed for hidden layers and opened the door for the development of deep learning. But still they were only regarded as an interesting concept with limited real-world application.

By the early 2000s I was creating my first AI network at university; facial recognition between men and women. I thought my thousand data points was huge, but it only resulted in about 80% accuracy. Pretty much the same as the humans did with the same pictures. This project is also useful to describe the difference between "classical software" and "AI". If I used classical software methods it would be the developers' role to tell the software what to look for. In AI however all I do is program it how to learn from the data I give it. As a developer, I do not know how the AI tells the difference between a male and a female.

The leap which made the big difference to AI was big data and the ability to process it. We could now train AI using data sets of millions rather than hundreds. This improved its accuracy, allowed us to train deeper networks and solve more complicated, real-world problems. AI became a practical solution rather than a solution looking for a problem.

So it’s not a new subject. AI has been on a long journey to get to this useful point. It’s had winter periods where nothing much happened, but we now seem to be at the beginning of a new stage.

Today, big data has given AI the references it needs and programming gives it the ability to acquire and apply knowledge and skills. Over a hundred years after we first started to understand neural networks, AI is now ready to solve real problems.

If you enjoyed this article please subscribe to ckx as we will be releasing similar content aimed at demystifying AI.