I’m about 100 pages from the end of Steven Pinker’s book How The Mind Works (previously blogged on the subject of evolution). While I have to grant the man a lot of credit for his writing style — funny, clear, often insightful — I’m getting the same feeling of vague creepiness that reading some of Richard Posner’s books (e.g., The Economics of Justice) gives me. It is the feeling that standing before me is a True Believer, who is trying to paint the entire world using a brush intended for housepainting. He has precisely one tool at his disposal: in Pinker’s case it is the theory of evolution, and in Posner’s it’s the hypothesis that what is good for society is a policy of wealth-maximization. There is substantial overlap between them; the evolutionary rationality that Pinker describes can compute probabilities and behave in a way that maximizes various expected values. In neither case — Pinker’s or Posner’s — is this a voluntary process, necessarily: the brain hasn’t been wired to allow us to compute these probabilities at a conscious level, but we behave as though it has.
But the world gets complicated: there’s more than one person in the world. I may be trying to maximize the probability of passing along offspring, but so are you. So are the other 6.3 billion people alive right now. So are all the world’s animals and vegetables. Natural selection is an enormous game (in the technical sense) played out over millions of years.
In the midst of this complexity, I hope I’m forgiven for some skepticism about the explanation for even very low-level processes, like the design of the human eye. Higher-level processes, like mate selection? Forget it. I can accept that one of the factors feeding into marriage is that we’ve been genetically programmed to find those people who maximize the probability of our genes carrying on for another generation. But there’s a lot left to explain. Pinker discusses the special bond we have with our biological children, and the lack of a bond between us and stepchildren. He points to a fair bit of evidence that the “evil stepmother” fable shows up across cultures. Fair enough. But how does he explain all the world’s good stepparents? How does he explain rare events that seem to serve no evolutionary purpose? Pinker’s “model” (I hesitate to call it that, because it seems so free of actual mathematics) predicts that altruism toward one’s kin will be much stronger than altruism toward strangers. For genetic reasons that makes sense; your family is much more likely to help you than strangers are, hence devotion to your family is likely to pass your genes on to a new generation. But then how does a Mother Teresa come about? We’ve had millions of years of evolution; why wouldn’t she have been selected against millennia ago?
It’s the exceptions that trouble me. If there were any mathematics behind these models, I imagine there would be a straight line superimposed over a widely-scattered random spray of points. The noise seems as though it would overwhelm the signal. Pinker shows us the signal and leaves out the noise.
Then there’s the matter of biological plausibility, which Pinker seems to come within a hair’s breadth of violating at all times. Gigerenzer et al.’s book Simple Heuristics That Make Us Smart argues pretty convincingly that strong assumptions about the Bayesian makeup of the brain simply couldn’t be how we operate: the computations are far too complex (computing posterior probabilities only became feasible with the advent of computer simulation), and we often have to make very quick decisions that a Bayesian brain wouldn’t have time to do.
Pinker comes close to proposing a Bayesian model for eyesight. A ray-tracing program that predicts how light would look after bouncing off a complex surface is solving a problem that’s inverse to what the brain has to do: the brain must ask, “What must the world look like that would create the pattern of light I have in front of me?” This problem is unsolvable as stated: multiple worlds could create a particular light pattern. So then How The Mind Works goes into a fantastic, lengthy, and super-interesting discussion of the sort of assumptions that the brain must append to the world: assumptions about the world’s essential regularity. Various optical illusions trick us into seeing something that isn’t there, precisely by exploiting assumptions about the world that don’t hold. Having reduced the space of possible models down to something manageable, the mind — Pinker suggests — chooses amongst the models by picking the one that maximizes a Bayesian likelihood ratio:
How does our 3-D line analyzer use Bayes’ theorem? It puts its money on the object that has the greatest likelihood of producing those lines if it were really in the scene, and that has a good chance of being in scenes in general. It assumes, as Einstein once said about God, that the world is subtle but not malicious.
But again, this is very very hard. And it’s not even clear that the problem is well-specified, even after eliminating some models that don’t meet regularity assumptions. Pinker mentions this problem elsewhere in the book: if I want to figure out the probability that I’ll get cancer, should I look at the set of all males born in Vermont in 1978 who don’t smoke, who weigh 155 pounds, who have been at one time or another unemployed, who graduated from college, and who have white middle-class parents? Make the reference class too small, and I’ll be its only member. Make it too large, and the probability you get out the other end will be a really bad approximation. So which reference class is your brain supposed to be picking?
This could then turn into a long discussion of how the brain might assign probabilities to events. Gigerenzer et al. make a good case that simple assignments of probability — those that don’t, in fact, use all the available information, and that don’t use statistical models with certain optimality properties (like least squares, or the Bayesian approach) are better in the wild: they’re more “ecologically rational,” to use Gigerenzer et al.’s phrase. They show in a wide range of datasets that simple rules generalize more easily than complicated statistical procedures like neural networks or least squares. We could use a more complicated procedure like the Akaike Information Criterion that chooses the number of parameters in the model to maximize generalizability, but then we get even further from a plausible model of what the brain is actually doing.
To Pinker’s credit, he is an experimentalist, and the beginning of his book is devoted to the great modeling power that computers bring us. Want to hypothesize about human nature? We don’t need vapid theorizing anymore: we can simulate a mind on a computer — the blankest of slates — and build in all the assumptions that our brain is supposed to adhere to. The trouble is that How The Mind Works only uses computer models a little bit — to show us how they might work in specific circumstances (say, figuring out where the light on a fanfold object is coming from). Everywhere that these models might cast some doubt on his own work, they’re mysteriously absent from Pinker’s book.
My point here is twofold. One is that Gigerenzer et al. have made me skeptical of any very complicated model of how the brain works. The second point is that any part of this discussion is enormously complicated and would require a book-length treatment of its own; Simple Heuristics That Make Us Smart is one such treatment, and it only touches on a small part of what Pinker discusses.
Pinker and Posner can afford to talk this generally about their subjects, because there’s no peer review to call them on their individual claims; peer review for a book is much different from peer review for a scholarly article, and both authors’ books count more as polemics. I worry that by skimping on this much detail for subjects that demand it, both authors will have convinced the world that a particular simple dogma describes it well, without really giving credit to all the nuanced arguments against them.
