Quale — a new word I will almost certainly never use.
It does however represent a concept we all have to wrangle with. Forget statistical models, literally no representation is complete. I tell you to imagine a red house – you imagine it. But was it red? Or maroon, burgundy, pink, or orangish? Not just a matter of precision, what we are communicating depends on what we both think of as ‘red’, or ‘maroon’, or ‘burgundy’, or whatever else. We might also have ideas about what sorts of red are ‘house’ appropriate. An upper-level ontology might suggest to us red-ness that is universal. But no houses in my neighbourhood are the bright lego red? Why not?
Some of what Schade writes reminds me of Intro Stats: error being present in every single observation. This sort of error can be thought of as explained and unexplained variance. Variance is present in all data; the unexplained variety being what may have risen due not only apparatus error, but also what we describe as uncertainty in data.
Schades temperature example is handy: the thermometer doesn’t read 14 degrees – it reads 13.5-14.5 with 68% probability. The stories we tell aren’t about what we say, but what we mean. This sort of anti-reductionism is also at the root of complexity theory. Acknowledging that we cannot characterize systems as static and linear components disregards the emergence of something that can explain why complex things are greater than the sum of their parts. Applied machine learning research also appreciates this anti-reductionism: the link to AI Schade makes, I THINK is about how applied machine learning researchers aren’t really interested in determining the underlying relationships of phenomena – only the observed patterns of associations. Methods that neglect the former and embrace the latter perspective explicitly consider their data to be incomplete and uncertain to some degree. Though to be honest this connection seems forced in the paper, but I’m happy to help force it along. 🙂