Interview excerpt (YouTube) in which Noam Chomsky critiques deep learning for language: he acknowledges engineering usefulness (e.g., Google translation/parsing) but argues it offers no scientific insight into human language. He emphasizes that DL finds patterns from huge corpora but can succeed on systems that violate core properties of natural language (e.g., structure dependence), so corpus-driven performance does not imply understanding.
Chapters
Let me ask you about a field of machine learning, deep learning. There's been a lot of progress in neural network based machine learning in the recent decade. Of course, neural network research goes back many decades. What do you think are the limits of deep learning of neural network based machine learning?
Well, to give a real answer to that, you'd have to understand the exact processes that are taking place, and those are pretty opaque. So it's pretty hard to prove a theorem about what can be done and what can't be done. But I think it's reasonably clear. I mean, putting technicalities aside, what deep learning is doing is taking huge numbers of examples and finding some patterns. Okay, that could be interesting in some areas it is. But we have to ask here a certain question. Is it engineering or is it science? Engineering in the sense of just trying to build something that's useful or science in the sense that it's trying to understand something about elements of the world. So take, say, Google parser, we can ask that question, is it useful? Yeah, it's pretty useful. You know, I use a Google translator, so on engineering grounds, it's kind of worth having like a bulldozer. Does it tell you anything about human language? Zero. Nothing. And in fact, it's very striking from the very beginning. It's just totally remote from science. So what is a Google parser doing? It's taking an enormous text, let's say the Wall Street Journal corpus, and asking how close can we come to getting the right description of every sentence in the corpus? Well, every sentence in the corpus is essentially an experiment. Each sentence that you produce is an experiment which says, am I a grammatical sentence? The answer is usually yes. So most of the stuff in the corpus is grammatical sentences. But now ask yourself, is there any science which takes random experiments which are carried out for no reason whatsoever, and tries to find out something from them? Like if you're, say a chemistry PhD student, you want to get a thesis, can you say, well, I'm just going to do a lot of, mix a lot of things together, no purpose, just, and maybe I'll find something. It'd be left out of the department. Science tries to find critical experiments, ones that answer some theoretical question, doesn't care about coverage of millions of experiments. So it just begins by being very remote from science. And it continues like that. So the usual question that's asked about say, a Google parser is how well does it do? Or some parser, how well does it do on a corpus? But there's another question that's never asked. How well, does it do on something that violates all the rules of language? So, for example, take the structure dependence case that I mentioned. Suppose there was a language in which you used linear proximity as the mode of interpretation. These. Deep learning would work very easily on that, much more easily than an actual language. Is that a success? No, that's a failure. From a scientific point of view, it's a failure. It shows that we're not discovering the nature of the system at all because it does just as well or even better on things that violate the structure of the system, and it goes on from there. It's not an argument against doing it. It is useful to have devices like this. So.
Yes. So neural networks are kind of approximators that. Look, there's echoes of the behavioral debates, right? Behavioralism more than echoes.
Many of the people in deep learning say they vindicated. Terry Sanjoski, for example, in his recent book, says this vindicates Skinnerian behaviors. It doesn't have anything to do with it.
Yes, but I think there's something actually fundamentally different when the data set is huge. But your point is extremely well taken. But do you think we can learn, approximate, that interesting complex structure of language with neural networks that will somehow help us understand the science?
It's possible. I mean, you find patterns that you hadn't noticed, let's say, could be. In fact, it's very much like kind of linguistics that's done what's called corpus linguistics, when you suppose you have some language where all the speakers have died out, but you have records. So you just look at the records and see what you can figure out from that. It's much better than. It's much better to have actual speakers where you can do critical experiments, but if they're all dead, you can't do them. So you have to try to see what you can find out from just looking at the data that's around. You can learn things. Actually, paleoanthropology is very much like that. You can't do a critical experiment on what happened 2 million years ago. So you're kind of forced just to take what data is around and see what you can figure out from it. Okay. It's a serious study, Sam.
