Q&A: Neural Networks and Intelligence
Neural Networks and Intelligence
Question
Hello Rabbi,
I recently read your article, “What Is Intelligence? On Free Will and Judgment (Column 35)” (https://mikyab.net/posts/1819#_ftnref1), in which you argued that since a computer has no judgment, it cannot be considered intelligent. You noted there that this is also true of a neural network; there too, everything is deterministic.
And yet, it seems important to me to distinguish between “classical” software and a neural network.
If I want to perform a complicated calculation, for example 36786489×46891388, I can do it on my own, but it would require a lot of effort, and there is a good chance I would make a mistake. So it is preferable for a computer to do the calculation, when I know at every stage exactly what the computer will do. Similarly, if I want to calculate the area under a graph (a certain integral), I can write software that divides the x-axis into very small parts (dx), calculate the value of the function in each part, and approximately obtain the area (numerical calculation). Here too, theoretically I myself could calculate the area that way, but it would be very difficult for me, and very inaccurate.
A neural network works in a very different way. You need to train the software. For example, you can teach a program the rules of chess and have it play against itself millions of times, and that way it learns and improves until it reaches a level higher than any human being.
Here the programmer only needs to know the rules of chess, and nothing beyond that. The programmer cannot explain what the program is doing, because he himself does not understand! So how can one say in such a situation that the intelligence comes from the programmer?! And if it does not come from the software itself, then where does it come from?
Answer
I did not understand the claim. Clearly, in a neural network too, the intelligence is that of the programmer and of the developers of the theory of neural computation. Your ability to calculate includes the possibility of calculating באמצעות computational tools. Of course, if the operation of the computational tool results from tools created by developers and not only from you, then the intelligence of the computational tool comes from all of you together.
Discussion on Answer
You assume that the computer’s “intelligence” comes from somewhere. How do you know that? It can be created out of nothing (that is, from training done by an ordinary person, so long as there is a wise person who figured out that neural networks can be trained).
By the way, I explained in that column that the “intelligence” of water or of birds also comes out of nothing in that sense.
Let us say a company sells very user-friendly neural networks. The network does not know how to do anything, but it can easily be taught. All you have to do is teach it the rules of whatever game you want, and then the network will play against itself a million times until it becomes expert at it.
I buy such a network for myself. Right now, within the closed system that includes me and the network, there is no experience or “intelligence” with respect to the game of chess. Now I teach it the rules of the game, press the button, and wait a short time. Now within that same closed system, there is a creature that can beat any person in the world at chess. I want to argue that the computer has some kind of intelligence, because otherwise this knowledge was acquired out of nothing