Why are brains so much better at learning/improvising than computers?
In: 7
Computers are called that way because they can compute *anything*. They are universally programmable for any task. The trouble is that it’s not easy to figure out how to program them to do really complicated things.
Now, when I say “complicated”, you might think about things like math or chess. But those are actually simple from the point of view of a computer, because they follow very clearly defined rules. You can write all the rules of chess on a small piece of paper, and it’s not that hard to explain them. But how do you explain, completely and precisely, how to identify a face in an image, based on its pixel values? This is something that we find so easy that we don’t even think about it, but turns out to be really difficult to express in a computer program.
What we have come to realize more and more in recent years is that this is not the way to go. It’s much better to write a program that allows the computer to learn on its own. Then you give it a bunch of images, and you say “these ones contain faces, these ones don’t – please learn a way to decide which is which”. And you leave it to the algorithm to come up with a good way to do it.
These algorithms work with *artificial neural networks*. These are simulations of computational units that are simplified versions of the neurons in our brains. Neurons connect to other neurons, and what they compute depends on which neurons they are connected to and how strongly. This means that you can change what’s being computed by altering these connection strengths, which turns out to be very helpful for learning things, as you can keep making small changes to the connections and see whether things improve or not (and in fact, you can use maths to figure out directions in which to change the connections that are likely to lead to improvement, so that you don’t just have to try at random).
The brain itself is also a neural network, only a much more complicated one. And the brain has its own learning algorithm(s) that allow it to adapt the connections between neurons to learn new things. We don’t know exactly how these learning algorithms work – this is a hot topic of research.
So in short, the brain comes pre-wired as a learning machine. Computers do not, but they can be programmed to learn, and when you do, they can start do to things similar to what brains can do.
When it comes to improvisation, that’s more a question of having enough world knowledge to be able to adapt to different scenario’s you haven’t seen before or haven’t specifically been trained for. Most AIs struggle with this as they are trained on rather specific tasks, so they can do those tasks very well but not anything new that you give them. Recently of course this is starting to change to with systems like ChatGPT that can answer questions about many different topics (but still can’t, for instance, ride a bike, or recognize images, as that’s not what these *large language models* are trained for). The more general world knowledge an AI has learned, the more it will be able to improvise.
Latest Answers