Learning about Machine Learning Through a Toddler’s Eyes
Nothing amazes me more than the wonder of our existence. And nothing ever cried louder about that wonder than having become a parent. First it’s the whole marvel of generating a new living being. But later, another wow appears: the way that this brand new living being learns.
A technological mind that I am, it’s inevitable to see parallels between a) how he (somewhat mysteriously) learns to identify items in the world around him and b) Machine Learning—despite how superficial my knowledge about the latter is. Let me take you through his journey (and ours, as parents) of learning.
We live in an apartment overlooking a moderately busy road. Our apartment’s windows have a clear glass panel with a low windowsill below the window itself, which means he has had the opportunity to watch cars go by since a really early age.
We would thus often use that to our advantage as a way to entertain him, especially on rainy days (uninviting for going out). We’d look outside with him, and say “car” whenever a car passed. Needless to say, that became one of his first words, and he was soon the one uttering “car” with his finger pointing whenever a vehicle passed.
He was saying “car” for any vehicle, including the second most popular type in the road in front of us: buses. We then raised the bar: we started correcting him. When he pointed to a bus saying “car”, we’d say “no, that’s a bus”. Soon enough, he started calling cars “car” and buses “bus”.
When walking around the neighbourhood with him, we noticed he consistently got cars and buses right. But what about other types of vehicles? Well, his brain seemed to try its best to assign them to either of these two categories: anything as big as a station wagon or bigger deserved being labeled a “bus”. Fair enough—we never explicitly told him what made a car a “car” and a bus a “bus”, why would he know that a truck is neither?
And so it continued: we kept correcting him. “No, that’s a truck”, “no, that’s a van”. Again, soon enough, he was getting it right himself. He correctly identifies a bicycle, a motorcycle, a car, a van, a truck, a bus—without any of us ever having explained to him what tells them apart.
What fascinates me is that we’re not talking about children’s book-type oversimplified drawings of these vehicles—rather real ones, with a variety of sizes, shapes, and colours, and he’ll correctly recognise them even when facing a significantly odd one. For instance, when we showed him pictures of London buses for the first time (yay for Paddington Bear stories!), he immediately called out “bus”—despite the fact that they were double-decker and red made them significantly different from anything he’s used to encounter in Lisbon.
The point here is not to say my kid is an above-average genius, but rather to marvel at the power of the human brain. You might say “sure, but this is what happens with Machine Learning, it gets better the more data you give it”. True, but most Machine Learning systems out there only do that after someone has codified what to look for in these data. Meanwhile, this intelligent living being somehow found his way to his model only by seeing samples and hearing labels. What features are in this model? Hell if I know! It’s probably an intuition-based mix of size, proportions, shape, number of wheel axes, distance between the latter, windows, etc. — but we didn’t point out to any of these to him. Unlike a Machine Learning-based system, we didn’t program the model—or did we?…