As many of you may know, I have been thinking about the following problem for a while:
Suppose that scientists discover a high level property G that is prima facie multiply realized by two sets of lower level properties, F1, F2, …, Fn, and F*1, F*2, …, F*m. One response would be to take this situation at face value and conclude that G is in fact so multiply realized. A second response, however, would be to eliminate the property G and instead hypothesize subtypes of G, G1 and G2, and say that G1 is uniquely realized by F1, F2, …, Fn, and that G2 is uniquely realized by F*1, F*2, …, F*m. This second response would eliminate a multiply realized property in favor of two uniquely realized properties A third possible scientific strategy would be to keep G and add subtypes G1 and G2. What do scientists actually do?
With Carl Gillett, I’ve been arguing, in essence, that scientists take door #1. This answer is defended in a forthcoming paper with Carl, “The Autonomy of Psychology in the Age of Neuroscience” ,
In
Illari, P.M., Russo, F., and Williamson, J. Causality in the
Sciences. Oxford University
Press.
It has recently come to my attention that Michael Esfeld, Christian Sachse, and Patrice Soom have been developing views (roughly) along the lines of door #3. See, for example,
“Theory Reduction by Means of Functional Sub-Types”
Reductionism in the Philosophy of Science.
Esfeld and Sachse also have a book forthcoming from Routledge developing a version of door #3.
I’m reading through their stuff now, but the most jarring thing for me is that they claim that they are exploring subtyping as a logically possible thing for scientists to do. But, if this is merely a logically possible thing for them to do, and not something they actually do, then why think this has much to do with science or reduction?
All that’s rough, but take this post as a trailer for this topic.
Ken, thanks for the interesting post. If we bracket a bunch of issues pertaining to levels and properties, I’d say what door scientists choose depends on the situation. Some properties are eliminated (e.g., “jade”), others are split into subkinds (e.g., memory), and yet others are retained in the face of multible realizations.
To explain why scientists choose any particular door would take a lot of work, but roughly speaking, they eliminate properties that turn out to be “superficial” (i.e., having few if any characteristic causal powers?), split properties that turn out to have subkinds with interestingly different causal powers, and retain properties whose realizers have important causal powers in common.
BTW, I’m planning to read some of Esfeld and Sachse’s work in the near future, so if you have any more detailed comments on that I’ll be interested.
The view that Carl and I have is that what scientists do does depend on the situation. A bit more specifically, it depends on the “scientific needs” of the higher level theory and not merely on the discovery of differences in lower level realizers. In fact, contrary to the suggestion in the description of the strategy, it may be that the lower level theory has nothing to do with whether or not to split. Splitting is entirely determined by the higher level theory. We don’t say that in the paper, but this stronger conclusion may be the right one. This runs counter, also, I think, to things that Craver at least suggests in his paper on dissociable realization.
Whether I agree with what you say in the second paragraph really depends a lot on how you cash out “superficial” and “interestingly different”. So, in color vision, the human lens yellows with age. But, vision scientists do not subtype on the basis of different degrees of yellowing. Instead, the different degrees typically are counted as bases of individual variation. So, one might say that the differences in yellowing are not so superficial as to not be worth a vision scientist’s time to study. Nor, one might say, are they so uninteresting that they are not worth studying. But, going on, one might also say that they are not so deep and not so interesting as to form the basis for different subtypes of normal color vision.
I have just been looking over the Esfeld-Sachse-Soome stuff for about a week, but I will definitely be writing something up in the next few months. It’s just a matter of timing. I’ll send a draft along when I have one. Thanks for your interest.
Doors 1 and 3 are both clearly important.
When subtypes are functionally different, then door 3 is used. The property of being a photoreceptor is instantiated by (at least) four different subtypes. These four subtypes have obvious important differences.
The property of being an NMDA receptor subunit is instantiated by multiple subtypes that also confer different properties on cells, and have different rules of combinations with other subtypes.
On the other hand, two NMDA channel subunit types with a neutral amino acid change between might confer no changes to the subtype. In that case, they’d be referred to as the same subtype.
So I think I’m agreeing with you and Guiltiero. When there is no interesting functional difference at a finer grain, lump ’em togeter (door 1). When there are functional differences among subtypes, keep ’em split (door 3).
I think that’s what you guys were saying.
The lens yellowing case is interesting, but it isn’t clear this is door 1 or door 3. Babies’ eyes are usually blue/gray when they are born, and then turn into the adult color sometime in the first year. It’s the same object taking on a different property, so I’m not sure if that’s door one or door number 3. It seems you may need a different set of doors for such dynamic properties.
“So I think I’m agreeing with you and Guiltiero. When there is no interesting functional difference at a finer grain, lump ’em togeter (door 1). When there are functional differences among subtypes, keep ’em split (door 3).”
Well, I think things are more complicated than I have let on in the post (as opposed to the paper). So, the picture that Gualtiero and you seem to have is that small higher level differences doesn’t lead to splitting, where large higher level differences leads to splitting. The paper implicitly argues that the situation is often more complex.
The humans crystalline lens (a lovely example that I have just come upon this summer) has two features vision scientists care about: call them “blue light filtering” and “refractive power”. The simple picture just described suggests that large variations in refractive power will have large differences in near-sightedness or far-sightedness, so that we should split normal human color vision along the lines of these differences degrees of near-sightedness or far-sightedness. Those are functionally significant differences. But, what I think happens is that normalcy of color discrimination and normalcy of focusing ability are two distinct “dimensions” of human vision. So, scientists will not want to let large variations in color discrimination influence the taxonomy of focusing ability. Similarly, scientists will not want to let large variation in focal abilities to infect their taxonomies of color discrimination.
Maybe the most apt philosophical apparatus for explicating this idea is that of determinables and determinates. There are two determinables in human vision: color discrimination and focusing abilities. One does not want to create mongrel categories that combine these two.
This stuff makes an appearance near the end of the Aizawa-Gillett paper, but is not developed very much.