In: ILLC Magazine (13 december 2009)

Jelle Zuidema

Sometimes it is research that you completely disagree with that inspires you the most, especially if it is close to what you worked on yourself. In 2001, the journal Science published a paper titled “Evolution of Universal Grammar”, by mathematical biologists Martin Nowak and Natalia Komarova, together with computer scientist and linguist Partha Niyogi. The paper concerned mathematical models of language learning and evolution and showed, the authors claimed, that there must be quite detailed, innate knowledge of language for successful communication to be possible at all in a population. It strongly supported the nativist camp in the big debate in linguistics about whether and to what extent language is innate.

The math in the paper is very elegant – and it was a lot of fun to play around again with differential equations and bifurcations – but the problem with it, I found, is that the model is completely wrong. In the year following its publication, I spent a lot of time and energy in understanding where exactly it went awry. It is interesting to see how many people uncritically accept conclusions from papers with lots of math and the right rhetoric, even if very few of them, I am convinced, have actually bothered to go through the derivations.

For me, two earlier inspirations were crucial to discover the error in the Science paper. The first was the research of my MSc advisor, theoretical biologist Paulien Hogeweg. She always emphasized that in every model, implicit assumptions are made, and advocated a “multi-modelling” approach where one tries to design multiple models of the same empirical phenomenon. By comparing the behaviour of these different models, you often find surprising differences and discover hidden assumptions that you might want to reconsider. I think this is still an important message in cognitive science and linguistics, where too often researchers are too much in love with their own little models and fail to see the problematic assumptions hidden behind fancy notation.

The second inspiration was the work of my later PhD advisor, linguist Simon Kirby. He studied the first “iterated learning” models. His work helped me realize that language learning is a very special kind of learning problem, because the target of learning is not God-given, so to speak, but the result of the learning that occurred in earlier generations. That implies that the language that children need to learn reflects the learning biases of earlier generations of learners. This point may seem quite trivial, but it turns out that much of the formal work in learnability theory and many of the verbal arguments for the “poverty of stimulus” or “critical period” are put on their heads when you realize what it really means. By building a computational iterated learning model that closely resembled the mathematical model from the Science paper, and closely analyzing the quite different outcomes, I figured out that the error in the original paper was that it assumed a wrong upper bound.

I have moved on, of course – after spending perhaps a bit too much time on the nitty-gritty details of computational and mathematical models that few people really care about. One inspiration in the last few years stands out as a motivation for me to sometimes look up from such obscure modelling and consider the big questions instead, and that is Jared Diamond’s book Guns, Germs, and Steel , in which he describes how the enormous differences in power and technology between people on earth have come about since Homo Sapiens emerged in Africa. I have tried to make everybody I know read this book. It might be wrong in many details, but the overarching story is totally convincing to me; it showed me that sometimes it is research that you completely agree with that inspires you the most – but only if it is far beyond what you have worked on yourself.

References:

[1] J. Diamond. Guns, Germs, and Steel: The Fates of Human Societies , W.W. Norton & Company, 1997.

[2] S. Kirby. “Spontaneous Evolution of Linguistic Structure: An Iterated Learning Model of The Emergence of Regularity and Irregularity”, IEEE Transactions on Evolutionary Computation 5 (2): 102–110, 2001.

[3] J.D. van der Laan, L. Lhotka, and P. Hogeweg. “Sequential Predation: A Multi-Model Study”, Journal of Theoretical Biology 174 : 149-167, 1995.

[4] M.A. Nowak, N. Komarova, and P. Niyogi. “Evolution of Universal Grammar”, Science 291 : 114-118, 2001. Science 291: 114-118, 2001.