If you can teach a machine to perform a human task faster, more accurately and more consistently than a human, you’ve just unlocked machine learning. Not all human tasks can be done by machines.
Computer scientist and entrepreneur Andrew Ng has a general rule of thumb to understand what you can reasonably expect AI to accomplish: he says that any task that a human can do with less than 1 second of thought is a likely candidate for machine learning.
You see a person. You can tell whether they are smiling or not. Definitely takes less than a second. How do we teach a machine to do the same?
A traditional programming approach for a task like this would be to codify what a “smile” is, in a way the machine can understand.
This might involve first narrowing down the inspection to the lower half of the face, identifying the mouth, isolating the lips, applying a number of rules to see if the lips are curved upwards, the different degrees of curvature to include all the different ways people smile, from a just barely-there hint of a smile to a wide grin, from lopsided embarrassment to a self-satisfied smirk… And now are you thinking is this even possible? How do you codify all the nuances, the near-infinite number of attributes of a smile, and more importantly, of a non-smile?
You can’t. Not with a traditional approach.
With machine learning, it is certainly possible; but not with one glance and definitely not in less than a second. You would need hundreds of thousands of pictures of people, smiling and not smiling, and you would need a human to “teach” the program to output “smile” or “not smile”, as the case may be. Essentially, you are training a piece of software to do what human eyes and brains do – in this case, look and tell. The difference is that humans may only need to look once, and can intuitively tell, whereas machines need to look at lots of data to learn what humans can do in less than a second. Given enough good data, and robust training, the machine can eventually process a great number of images in one second. This is beyond the capacity of any human.
Once a certain level of accuracy is reached, and this is where we come to the mind-blowing bit – with enough data, and just data, the AI continues to learn on its own. It trains its network of neurons as it gets fed more data. It takes decisions. It creates insights. With just data. That’s all.
That’s all? Well, not exactly. It’s all good as long as the data is good. But when the data is of dubious quality, or of little relevance, or just not organized in a meaningful way, it can, at best, be worthless, or at worst, give you very wrong results. And this is where the sheen of AI starts wearing off. Perhaps you’ve driven into a dead-end street while trusting the map app more than your own judgment. Perhaps it rained when the AI weather model predicted clear skies. Perhaps the X-ray screening software couldn’t tell a hairline fracture from a scratch on the film. Perhaps a human did a better job of understanding a customer complaint than a chatbot.
Not so mind-blowing now, is it? Not to despair, however – we can still make AI work for us, if we do it intelligently enough. This involves precisely defining the task we want automated, in other words, keep the scope narrow. Think of a face recognition program that does only that – recognize faces from images. It does not in any way assume where they are from, or predict what they are going to do. Think of a chatbot that can only answer users’ queries on product information. It does not in any way attempt to recommend one product over another. Think of a model that does just one thing, and does it well, and does it faster and more accurately than a human, and you’ve got a winner. Sure, data is key, but even more crucial is a clear and specific idea of what the AI needs to do – and that depends entirely on a human.