>
> —["Crossing the Line" (extended lyrics)](https://www.youtube.com/watch?v=QbaMiK9oqJE), _Rapunzel's Tangled Adventure_
-Emily Cibelli, Yang Xu, Joseph L. Austerweil, Thomas L. Griffiths, and Terry Regier's ["The Sapir–Whorf Hypothesis and Probabilistic Inference: Evidence From the Domain of Color"](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0158725) is a cool paper about how language affects how people remember colors! You would expect that the design of the eye and its colorspace to be human-universal (_modulo_ [colorblindness](https://en.wikipedia.org/wiki/Color_blindness) and [maybe some women with](https://en.wikipedia.org/wiki/Tetrachromacy#Humans) [both](https://en.wikipedia.org/wiki/OPN1MW) [kinds](https://en.wikipedia.org/wiki/OPN1MW2) of green opsin gene), but not all languages have the same set of color words. There are some regularities: [all languages have words for light and dark; if they have a third color word, then it's _red_; if there's a fourth, it'll cover green or yellow](https://en.wikipedia.org/wiki/Basic_Color_Terms)—but the details differ, as different languages [stumbled onto different conventions](/2020/Jan/book-review-the-origins-of-unfairness/). Do the color category conventions in one's native tongue affect how people think about color, in accordance with the famous [Sapir–Whorf hypothesis](https://en.wikipedia.org/wiki/Linguistic_relativity)? Maybe—but if so, how??
+Emily Cibelli, Yang Xu, Joseph L. Austerweil, Thomas L. Griffiths, and Terry Regier's ["The Sapir–Whorf Hypothesis and Probabilistic Inference: Evidence From the Domain of Color"](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0158725) is a cool paper about how language affects how people remember colors! You would expect the design of the eye and its colorspace to be human-universal (_modulo_ [colorblindness](https://en.wikipedia.org/wiki/Color_blindness) and [maybe some women with](https://en.wikipedia.org/wiki/Tetrachromacy#Humans) [both](https://en.wikipedia.org/wiki/OPN1MW) [kinds](https://en.wikipedia.org/wiki/OPN1MW2) of green opsin gene), but not all languages have the same set of color words. There are some regularities: [all languages have words for light and dark; if they have a third color word, then it's _red_; if there's a fourth, it'll cover green or yellow](https://en.wikipedia.org/wiki/Basic_Color_Terms)—but the details differ, as different languages [stumbled onto different conventions](/2020/Jan/book-review-the-origins-of-unfairness/). Do the color category conventions in one's native tongue affect how people think about color, in accordance with the famous [Sapir–Whorf hypothesis](https://en.wikipedia.org/wiki/Linguistic_relativity)? Maybe—but if so, how??
-Cibelli, Xu, _et al._ discuss an experiment where people are briefly shown a color, and then try to match it on a color wheel, either at the same time, or after a short delay. People aren't just not-perfect at this, but—particularly in the delayed condition—show a non-monotonic pattern of directional bias: colors just on the "blue" side of the green–blue boundary are remembered as being relatively more bluish than they really were, but _very similar_ colors on the "green" side of the boundary are remembered as being relatively more greenish than they really were. (Where what counts as "blue" and "green" was operationalized by asking the same subjects to rate colors on a "not at all blue (respectively green)" to "perfectly blue (respectively green)" scale.)
+Cibelli, Xu, _et al._ discuss an experiment where people are briefly shown a color, and then try to match it on a color wheel, either at the same time, or after a short delay. People aren't just not-perfect at this, but—particularly in the delayed condition—show a non-monotonic pattern of directional bias: colors just on the "blue" side of the green–blue boundary are remembered as being relatively more bluish than they really were, but _very similar_ colors on the "green" side of the boundary are remembered as being relatively more greenish than they really were. (Where what counts as "blue" and "green" was operationalized by asking the same subjects to rate colors on a "not at all" to "perfectly" blue/green scale.)
-How to explain this curious pattern of observations? The answer is—Bayesian reasoning! [(The answer is always Bayesian reasoning.)](https://www.lesswrong.com/posts/QrhAeKBkm2WsdRYao/searching-for-bayes-structure) Our authors propose a model in which a stimulus is encoded in the brain as both a fine-grained representation of what was actually seen (this-and-such color perception, with some noise/measurement-error), and as a category ("green"). Then reconstruction of the stimulus from the both the fine-grained representation and the category, will be biased towards the center of the category, with more bias when the fine-grained representation is more uncertain (as in the delayed condition).
+How to explain this curious pattern of observations? The answer is—Bayesian reasoning! [(The answer is always Bayesian reasoning.)](https://www.lesswrong.com/posts/QrhAeKBkm2WsdRYao/searching-for-bayes-structure) Our authors propose a model in which a stimulus is encoded in the brain as both a fine-grained representation of what was actually seen (this-and-such color perception, with some noise/measurement-error), and as a category ("green"). Then a reconstruction of the stimulus that uses both the fine-grained representation and the category, will be biased towards the center of the category, with more bias when the fine-grained representation is more uncertain (as in the delayed condition).
-The model gains further support from a similar "two-alternative forced-choice" experiment, where people try to tell the difference between the originally-displayed color and a distractor (rather than picking from a color wheel). English speakers are better distinguishing between an original and distractor on opposite sides of the green–blue boundary. But speakers of Berinmo (spoken in Papua New Guinea) and Himba (spoken in Namibia) don't have the green–blue distinction: the Berinmo _wor_–_nol_ and Himba _dumbu_–_burou_ boundaries fall between what English speakers would call _yellow_ and _green_. And as the model predicts, Berinmo and Himba speakers respectively do better at distinguishing between original and distractor on opposite sides of the _wor_–_nol_ and _dumbu_–_burou_ boundaries!
+The model gains further support from a similar "two-alternative forced-choice" experiment, where people try to tell the difference between the originally-displayed color and a distractor (rather than picking from a color wheel). English speakers are better distinguishing between an original and distractor on opposite sides of the green–blue boundary. Speakers of Berinmo (spoken in Papua New Guinea) and Himba (spoken in Namibia) don't have the green–blue distinction, but the Berinmo _wor_–_nol_ and Himba _dumbu_–_burou_ boundaries fall between what English speakers would call _yellow_ and _green_. And as the model predicts, Berinmo and Himba speakers respectively do better at distinguishing between original and distractor on opposite sides of the _wor_–_nol_ and _dumbu_–_burou_ boundaries!
-In addition to superior cross-category discrimination, the model also successfully predicts _within_-category effects. Suppose one stimulus, which we'll call _A_, is a more central example of its category than stimulus _B_. Then in the two-alternative forced-choice paradigm, it's easier to distinguish _A_ as an original from distractor _B_ than it is to distinguish _B_ as an original from distractor _A_, because the exceptional case _B_ [regresses towards the mean](https://en.wikipedia.org/wiki/Regression_toward_the_mean) in memory.
+In addition to superior cross-category discrimination, the model also successfully predicts a _within_-category bias. Suppose one stimulus, which we'll call _A_, is a more central example of its category than stimulus _B_. Then in the two-alternative forced-choice paradigm, it's easier to distinguish _A_ as an original from distractor _B_ than it is to distinguish _B_ as an original from distractor _A_, because the exceptional case _B_ [regresses towards the mean](https://en.wikipedia.org/wiki/Regression_toward_the_mean) in memory.
Regular readers of _The Scintillating But Ultimately Untrue Thought_ know where I'm going with this! Why do we care about the further question of what "gender" someone is, if we already have fine-grained perceptions of how the person looks and behaves? Because our brains [use category-membership as an input into predictions when our perceptions are uncertain](/2020/Apr/book-review-human-diversity/#everyday-base-rates).
-If categories influence judgement on tasks as simple as _remembering colors_, then on theoretical grounds, I would expect the effect of gender on perception of people to be _much_ worse (that is, larger), because people are much more complicated than colors. With colors, what you see is _basically_ what there is: if your memories or perception of 500-nanometer wavelength light get rounded off slightly bluewards or greenwards depending on how many color words are in your native language, that's kind of bad compared to what a well-designed AI with access to the pure, unmediated colorspace could perceive, but at least that bias is only acting on _one_ dimension. In contrast, your observations of a particular person are going to be much sparser than everything your brain might want to predict about that person, so your priors about what humans are like are going to have to do _more work_—work that relies on the ways ([some blurry](https://pubmed.ncbi.nlm.nih.gov/16173891/), [some unseen](/2020/Apr/peering-through-reverent-fingers/)) that female humans are different from male humans.
+If categories influence judgement on tasks as simple as _remembering colors_, then on theoretical grounds, I would expect the effect of gender on perception of people to be _much_ worse (that is, larger), because people are much more complicated than colors. With colors, what you see is _basically_ what there is: if your memories or perception of 500-nanometer wavelength light get rounded off slightly bluewards or greenwards depending on how many color words are in your native language, that's kind of bad compared to what a well-designed AI with access to the pure, unmediated colorspace could perceive, but at least that bias is only acting on the _one_ dimension of color. In contrast, your observations of a particular person are going to be much sparser than everything your brain might want to predict about that person, so your priors about what humans are like are going to have to do a greater share of the work—work that relies on the ways ([some blurry](https://pubmed.ncbi.nlm.nih.gov/16173891/), [some unseen](/2020/Apr/peering-through-reverent-fingers/)) that female humans are different from male humans.
Transgender people are in a uniquely epistemically privileged position to observe this process, as the change from not-passing to passing is simultaneously a _small_ one as far as the person themselves is concerned, and a _large_ one as far as how the person is percieved by others. In a couple paragraphs that make me feel sad and jealous (I can't say _dysphoric_ because I don't know what that word means), Julia Serano explains what it's like to cross that line (in Ch. 8, "Dismantling Cissexual Privilege", of _Whipping Girl: A Transsexual Woman on Sexism and the Scapegoating of Femininity_):
Serano goes on to suggest that social gender exists, not in the way individuals perform gender, but in how others perceive it, and that therefore efforts to create a less oppressive world must involve dismantling cissexist and heteronormative assumptions: "if we truly want to bring an end to all gender-based oppression, then we must begin by taking responsibility for our own perceptions and presumptions[; t]he most radical thing that any of us can do is to stop projecting our beliefs about gender onto other people's behaviors and bodies."
-I can see how one might derive that lesson from the described experiences of transitioning, but I think it's ultimately a flawed generalization from a _necessarily_ unrepresentative experience. Trans people are [adversarial examples](https://christophm.github.io/interpretable-ml-book/adversarial.html): the ways people treated Serano differently after she transitioned despite Serano being the same person the whole time, are not _arbitrary_: that happened _because_ the fact that Serano looked like a woman, prompted people to use mental models trained against the distribution of women. (There might be [_reasons_ going back hundreds of millions of years](https://en.wikipedia.org/wiki/Infanticide_in_primates) for primate mothers to become suspicious or fearful of males near their children.)
+I can see how one might derive that lesson from the described experiences of transitioning, but I think it's ultimately a flawed generalization from a _necessarily_ unrepresentative experience. The ways people treated Serano differently after she transitioned despite Serano being the same person the whole time, are not _arbitrary_: that happened _because_ the fact that Serano looked like a woman, prompted people to use mental models trained against the distribution of women. (There might be [_reasons_ going back hundreds of millions of years](https://en.wikipedia.org/wiki/Infanticide_in_primates) for primate mothers to become suspicious or fearful of males near their children.)
-The change in other agents' behaviors elicited by [sending the signals](https://www.lesswrong.com/posts/4hLcbXaqudM9wSeor/philosophy-in-the-darkest-timeline-basics-of-the-evolution) of a different type is so dramatic specifically [_because_ it's a rare, off-equilibrium play](https://www.lesswrong.com/posts/YptSN8riyXJjJ8Qp8/maybe-lying-can-t-exist): if there were much more probability-mass just on either side of a line people are using to make predictions and decisions, then _the line wouldn't be there_.
+In the same chapter of _Whipping Girl_, Serano mentions that in her days of identifying as a male crossdresser, she found it easier to pass in suburban areas rather than cities, "where people were presumably more aware of the existence of gender-variant people."
-If we truly want to bring an end to gender-based oppression, it seems like such an ambitious social engineering project would be _helped_ by understanding elementary points like this? The most _useful_ thing that any of us can do is to to start having our beliefs about gender reflect the [laws of inductive inference](https://www.lesswrong.com/posts/eY45uCCX7DdwJ4Jha/no-one-can-exempt-you-from-rationality-s-laws) as applied to the special case of other people's behaviors and bodies. Not to disagree with the expedition to [a land where the children are free](/2017/Dec/theres-a-land-that-i-see-or-the-spirit-of-intervention/), but for the fear that between here and there, there are so many boundaries. Far more than I know.
-[TODO: edit the ending to not be so desperate to riff off Serano or Tangled or wave my usual "Truth!!" flag, but just be really clear about the statistical point I'm making]
+The change in other agents' behaviors elicited by [sending the signals](https://www.lesswrong.com/posts/4hLcbXaqudM9wSeor/philosophy-in-the-darkest-timeline-basics-of-the-evolution) of a different type is so dramatic specifically [_because_ it's a rare, off-equilibrium play](https://www.lesswrong.com/posts/YptSN8riyXJjJ8Qp8/maybe-lying-can-t-exist).
-[TODO: add a study about Sapir–Whorf in languages with gramatical gender]
+https://www.lesswrong.com/posts/yLcuygFfMfrfK8KjF/mutual-information-and-density-in-thingspace
+
+If there were much more probability-mass just on either side of a line people are using to make predictions and decisions, then _the line wouldn't be there_.
projects / Ultimately_Untrue_Thought.git / commitdiff