On Patches and Patterns: Local Knowledge and Scientific Success

It’s often said that science strives towards generality, looking for laws and principles about reality that admit of no exceptions, or as few as possible. Some even go as far as saying that unity is a standard of scientific success, that an ideal scientific knowledge would be one simple, unifying, and universal theory of everything.

Others, as you’d expect, disagree. One angle of attack is to claim that there are plenty of branches of science that simply have no such thirst for generality. In vast swaths of the biological sciences, for example, the aim of the game isn’t finding general laws but about narrowing to a single domain and getting to know it as well as possibleーgrowth regulation in the fruit fly, or the nesting behaviour of the northern flicker. To put it another way, much of science happily confines itself to just one pocket of the world without (much) concern for what’s going on elsewhere.

What good is there in restricting attention in this way? Is it just because our resources (time, money, interest) are finite, and so studying reality a piece at a time is the best we can do? That must be part of the storyーwe are limited beings, after all. But there’s good reason to think that this local knowledge is more than just an unfortunate and temporary side-effect of human limitations. In fact, local knowledge has rich benefits of its own.

One example comes from Davis Baird and his paper Engineering Realities (2010). There he considers a problem faced in quality control in manufacturing, and an ingenious solution developed by the start-up company Ometric. When you’re making a consumable product like dog food or medicine, you have to check the composition of each batch to make sure it contains the right amount of everything. (In dog kibble, it turns out, water content is very important.) This is typically done by a technique called spectrometry: pass a beam of light through the sample, then split the emerging light into separate wavelengths using a prism and measure how much of each wavelength has got through. Different chemicals absorb different profiles of wavelengths, and more of a chemical absorbs more of those wavelengths; so spectrometry gives you information about what is in the sample, and how much. The quality control problem in manufacturing is one of efficiency: spectrometry is a delicate process owing to the precision required in splitting the light beam. It also has to be done apart from the production line: to test a batch, you have to take a sample aside and test it, which destroys the sample. If it fails the test, the whole batch is scrapped, even though some of it may be fine. This is bad for business.

Ometric’s solution is a clever one. It borrows a statistical technique called principal component analysis (PCA) from psychometrics, the measurement of psychological traits like intelligence. Psychometricians have figured out that you can distill the results of a variety of psychological tests into a few “principal components”, which contain almost all the information about the wider range of tests in an easy-to-digest form. (There is a lively debate about whether the principal components represent the “real” traits underlying the results, which each test measures only imperfectly.) Ometric borrowed PCA and applied it instead to the measurement of chemicals: just as in psychometrics, you can reduce a range of tests of different wavelength absorptions to just a few. Once you’ve done that, you can design a filter that keeps only those few important wavelengths, then pass the beam through the test sample. That way, you don’t have to split the beam and then measure the passage of each wavelength afterwardsーyou just measure the intensity of the whole beam. Using this technology, then, you can test every bit of material in-line, throwing away only the ones that don’t pass this sturdier test procedure.

Here’s the lesson: the reason you don’t have to use a whole spectrometer for quality control is that production lines are what Baird calls “engineered realities”ーpockets of order and stability in which much is possible that wouldn’t be outside it. Things can go wrong while making a material, but only so much: too much of this ingredient, too little of that, particular contaminants. This amounts to a small and manageable set of variables to measure and control. The filter in Ometric’s system is tailor-made with that knowledge in mind: it won’t work for a different material, but that’s not a problem. By knowing your environment, you can do things quicker, more efficiently, and with more precision.

Engineered realities are stable, predictable, and also of our own making. But the point about the value of local knowledge doesn’t just apply to environments of our own design: there are eddies in the chaos to be discovered as well as created, and knowing those can be just as useful. To give just one example, a great leap in the science of communication was to observe that natural languages have pattern and order: if you know that pattern, you can build that knowledge into the technology we use to talk to each other.

With that, we can return to the question we started with about science as local versus general knowledge. What exactly does this dispute come down to? Why, for instance, should we expect scientific advancement to either tend towards unity or diverge into patches? One way of interpreting this dispute is that it’s a disagreement about what the world is like. Some have the intuition that the world is fundamentally simple; since it’s science’s job to say what reality is fundamentally like, the thought goes, science itself should take simplicity as a mark of progress. Herbert Simon (1996) expresses a similar intuition: “The central task of a natural science is to make the wonderful commonplace: to show that complexity, correctly viewed, is only a mask for simplicity; to find pattern hidden in apparent chaos” (p. 1). Others doubt that there’s any fundamental underlying simplicity. Ken Waters (2016), for one, takes something like the latter view: for him, the most general thing we can say about reality is that there is no “general structure” to be found. The best we can do, in that case, is to get to know patches of it.

So looked at this way, it seems that this disagreement comes down to appearance versus reality: For some, chaos, complexity and disorder are an illusion, or simply emergent from a simple foundation. For others, reality itself is a mess, and the appearance of simplicity comes from, say, restricting our attention to pockets of that reality where there’s order to be found. So who’s right? Is simplicity the reality and chaos the illusion, or the other way round?

I think the answer may well be: neither. Case studies like Baird’s, I think, illustrate something important about science and how it succeeds. On the one hand, it demonstrates that a lot of knowledge is irreducibly local. Not only is that okay, but that local knowledge can be extremely useful for a lot of things that we, including scientists, do. On the other hand, Ometric’s borrowing of a technique developed for psychometrics shows that many of the tools we develop for one thing can be useful for many others. Crucially, noticing this requires the ability to see structural similarities among apparently very different thingsーin this case, measures of human characteristics and the analysis of chemicals. (This is an example of the value of abstractionーa theme for a later post.) The lesson, I think is this: scientific success depends equally on noticing similarities and differences; on abstraction and distinction alike. If this says anything about what the world is “really” like, it’s that similarity and difference are features of reality all the same: while it’s certainly patchy in one sense, there are similarities to be found between the patches, both within and across scales. Whether the similarities or differences are more important in a given case depends on what we’re trying to do. Science succeeds, it seems, when it figures out whether two jobs call for similar tools or different ones.

-Oliver

References

Baird, D. (2010). Engineering Realities. Spontaneous Generations: A Journal for the History and Philosophy of Science, 4(1), 94–110

Simon, H. (1996). The Sciences of the Artificial. The MIT Press.

Waters, C. K. (2016). No General Structure. In M. Slater & Z. Yudell (Eds.), Metaphysics in Philosophy of Science. Oxford: Oxford University Press