In self-driving car, this is what we do. Self-driving car uses a lot of sensors. But the one I'm familiar with is camera, it's the most complex one. They also have light sensors but they only detect a given frequency band. But In camera, we divide the parts of the different things in there by colour. Say it's the road, in front us. So here we're not playing with just one assumption, but to get started, let's make one. Now the camera can only takes video, or pictures, It cannot tell what's inside. Unless the programmer tell it to. The first assumption is say that the black or grey thingy in what the camera captures, is the road, therefore, the car is fine to go there. Now that's a good assumption, but what if there's a black car. The camera will still assume that it's a road, and it will think that it is actually an empty road when it's actually a car. So how can we avoid this? We Introduce another assumption, say that cars have a boxy shape. Or that streets in general is like a triangle, because from the camera perspective the road looks like this / | , there is a triangular thingy that represents the road. What else can we add? We can say, that, the sky is white, or blue. right. By assuming that, automatically we're also assuming that there is a contrast of colour between the road and the sky. One is dark and one is light. And that can also help us navigate. But what if it's raining, and the sky is dark. Of course we have to improve our algorithm, maybe look for other reference to distinguish the road from anything else. Maybe don't use the sky, but we could use the side of the road? Which could be green, because maybe we're surrounded by trees. But what if other cars are white, and the sky is white, the car could think that it's a dead end, because there is no more black thingy, no more road, when there is, and it's just been blocked by the white cars. In fact one of the most recent Tesla crash, is because their cars cannot distinguish white cars and snow or sky, i forgot which one exactly. But what tesla is very good at is software update. Just like our phone, tesla could update their car, on the move, giving it better performance, or better self-driving experience in a 5 minute software update. Of course it is limited to software improvement, and you can't program a honda to run as fast as a lamborghini, no matter how good the software is, it cannot be done. But maybe we can tailor its driving experience, fuel efficiency etc. The concept is very simple, but the algorithm in consist first, recognising the objects in the picture, in machine learning, we call this feature extraction. Not all objects are relevant, and this is an assumption too. If you detect a bird, that's not relevant is it. Compare to say detecting the streets. or the signs. The algorithm is what we call the model. We see everything as living in a simulation, so this AI lives in a virtual space. Just like ours, but in our world, object interaction are ruled by Maths and Physics. We have gravity, etc. Now in this AI virtual space, they too are ruled by some physics, not like ours, they don't have gravity etc. But in say self driving car, it is in their physics that green thingy where there a brown stick attached to that green thingy, that thing together, we call it a tree. And the algorithm gets better, the more trees it seen, right? There are many other types of trees, and the computers cannot recognise them all, so what we do is we abstract away from any type of trees, and we give the trees a general form, where it doesn't always return 100% correct detection that indeed it is a tree, but maybe a 70-80% correct detection on all trees, even if we never seen that trees before. So our goal is, given a sample of data, a sample of trees, we want our model, to be able to recognise, all the trees, including the ones that we've never seen. The sample data represents the assumption, the assumption that this is what trees looks like. Then we proceed to other objects, describing its shape in a mathematical form. I know, it's stupid. But that's what happening under the hood.
Man this is such long irrelevant explanation. But AI, and machine learning are just ridiculously difficult problems of probability.
