There’s one thing I don’t like about comparing DL with how a human brain works. That is, one thing apart from that the number of hurraoptimistic articles on DL reminds me eerily of hypes about perceptrons and Lisp machines, which were both soon followed by long AI winters.
That other thing is the assumption that (1) we can jump over all the simpler layers of our brains, those which were developed in order to respond to less abstract data with less abstract decisions, right to the place where we use DL to chat with us, or recognize tanks on satellite photos, or play highly symbolic board games. And in the same time we claim that (2) the neural networks we use are similar to real brains. I mean, I have nothing against (1). Computers playing Go better than Shuusaku? I jumped with joy when I learned that. But with (2) I think we should be more humble and more honest with ourselves.
DL does not create Artificial Intelligence which is more and more like our own. It creates AIs which obtain similar results, but in a very different way. They are much more focused on specific tasks, less general, less multi-purpose, and much more analytic. AlphaGo may be the best Go player in the world, but a human brain, which is only slightly worse, will probably know a number of other board games, how to drive a car, how to catch a tennis ball from the air, how to read and understand long texts in at least one human language (and base its decisions on that), and how to recognize not only tanks, but also millions of other objects.
And even if we say “OK, but we have AIs which do all these things, so we can build an AI which will be able to do them all together” it will still not be the same thing. Because for the real brain they are not different things — they are integrated. Placing a stone in Go is a result of a careful thought, but it is also a result of guiding a hand through space and putting an object recognized as a stone on a certain point on something recognized as a board. In the same time, the decision that motivated the move is based both on calculations and on metaphors, such as “now I will attack”, “I have to defend”, and so on, which hail both from instincts older than our ability to reason, and from our broader views on how to deal with the world.
Then, of course, we can say “OK, we will integrate these abilities, no biggie, just give us another few years”. And I can probably agree. It is possible. But the way we approach it is very different from how the real brain was developed in the first place, so the way the integrated AI will work will also be very different. For example, I think with this approach we will probably never get the instincts right. They’re just not important enough for practical purposes and in the same time they’re very difficult to grasp (what exactly is an instinct, anyway?).
Of course I write all this because I’m biased. I’m mainly interested in how AI can make very quick and good decisions based on insufficient, unreliable data. Like, for example, a bee flying over a field. Its brain is very small, it has to process data very fast, the data is vague, coming from flawed, organic sensory instruments, it has to watch for dangers and know how to avoid them, but in the same time it has to have some internal state, remembering its main tasks, where is the beehive, etc. How do you put all this in a bee? It’s something totally different from Deep Learning with its Big Data and giant machines. But I have a feeling that learning that will move us closer than DL towards the idea of how our brains work.
