Traditional artificial intelligence (AI) has come on in leaps and bounds from its earliest beginnings, but it did hit a roadblock. As fast as mainframe computers could run computations, they were only as smart as the programmers who accessed the data and coded how to look at it. The computer program – the AI – lacked the ability to know or truly understand the data it was seeing.
A new approach was needed to radically change how coding AI was approached. Only this way could computers get past their limitations. It wasn’t a question of them becoming smarter per se – they could already run more computations per second that the average human – but they were not sophisticated in understanding what they were seeing. In other words, while they could learn to play chess, they found the subtle nuances in the game harder to grasp; it’s not only about how many moves ahead you can play.
With machine learning, the idea is that computers learn by being fed information. As they see and experience more information, they get smarter. However, the same old problem exists. They must understand the information they’re being shown to appreciate exactly what it is and its relevance to what they already know.
For instance, it’s useful to recognize that a vehicle has driven past on the road. Only when being able to identify a vehicle with a driver inside – as opposed to it being some other object moving past the visual frame like a bird – can AI start to appreciate more about it. Once it knows it’s a car, if can measure speed and determine whether the vehicle is going faster than the permitted speed limit.
It may also draw a connection between the vehicle, it’s license plate and the record of the owner of that vehicle. Whilst these things can be captured independently, being able to pull the information together with the machine learning what’s relevant and what’s not is a powerful thing.
Computers and machine learning don’t function well when it cannot identify what something is. Seeing a photo, a human brain can identify different things in the image e.g. a person or a face, and recall who they are. It may recognize the Taj Mahal in the background from a previous image in a travel brochure and understand that this is probably a travel photo of someone on a trip through India.
These sorts of connections are only possible when images and other things can be broken down into their component parts. That requires humans to teach machines how to do that or to do it for them in many cases.
With data labeling, human workers look at information like a photo and break down the important factors about it. That may be the individual vehicles in the shot, the things of merit in a travel photo or the outline of a face to help with facial recognition.
Data labeling is a vital piece of the puzzle. Without being able to identify the most important things within a photo, the AI cannot spot them in a different photo or a live camera view. However, with enough different examples, a computer can begin to appreciate what different things mean. At that point, they’ve just taken a monumental leap forward in understanding.
AI is making leaps and bounds. The progress will only accelerate as computers learn to understand what they’re seeing and make decisions about it. It’s a bold new world.