Harnad, S. (2003b) Categorical Perception. Encyclopedia of Cognitive Science. Nature Publishing Group. Macmillan.
Differences can be perceived as gradual and quantitative, as with different shades of gray, or they can be perceived as more abrupt and qualitative, as with different colors. The first is called continuous perception and the second categorical perception. Categorical perception (CP) can be inborn or can be induced by learning. Formerly thought to be peculiar to speech and color perception, CP turns out to be far more general, and may be related to how the neural networks in our brains detect the features that allow us to sort the things in the world into their proper categories, "warping" perceived similarities and differences so as to compress some things into the same category and separate others into different categories.
Perez-Gay, F., Thériault, C., Gregory, M., Sabri, H., Harnad, S., & Rivas, D. (2017). How and why does category learning cause categorical perception? International Journal of Comparative Psychology, 30.
Perez-Gay, F., Thériault, C., Gregory, M., Sabri, H., Harnad, S., & Rivas, D. (2017). How and why does category learning cause categorical perception? International Journal of Comparative Psychology, 30.
Pullum, Geoffrey K. (1991). The Great Eskimo Vocabulary Hoax and other Irreverent Essays on the Study of Language. University of Chicago Press.
"CP occurs whenever perceived within-category differences are compressed and/or between-category differences are separated, relative to some baseline of comparison."
ReplyDeleteThis compressed within-category differences and separated between-category differences are intriguing, which makes me wonder how it happens and why it happens? The "how" question is answer later by the hypotheses provided by the computational and neural models of CP, that neural nets can selectively detect and emphasize some features (invariants) while ignore others (variants). In terms of why, is it determined by the environment where the organism lives? By evolution?
" "Whorf Hypothesis" (Whorf 1956; also called the "linguistic relativity" hypothesis), colors are determined by how our culture and language happens to subdivide the spectrum."
So we know that our categorical perception of colors has nothing to do with how we subdivide and name the color spectrum. On the contrary, this effect is probably an innate one, which was built in us by evolution. Therefore, can we say that in this case, the categorical terms we use for those categorical colors are not reasons but on the contrary, consequences? Is it that the language (at least sometimes) is a byproduct of categorization?
I'm not convinced that the neural nets response to the "how" question of categorical perception coming from computational models is an adequate answer (and that's agreed on in the text). On the topic of "why" that you mentioned, it seems like the answer is: it depends on what kind of CP you are talking about.
DeleteOne thing I noted while reading this second piece was the fact that, given what we know about visual CP, language grounding, and more, it seems possible that the "how" question's answer might also be: it depends on the type of CP. My prediction would be that the core processes involved in cognizing the in-group and out-group differences in objects/things/concepts/people might be localized in different regions of the brains, with different mechanisms that explain them. All in all, it seems like we might not be getting an answer to that anytime soon.
I agree that the "how" question has not yet been adequately answered. Particularly if we refer back to reverse engineering, do we know how the human brain categorizes and has this knowledge enabled for the building of that process?
DeleteSome categories are innate, but most are learned, based on corrective feedback from the consequences of doing the right or wrong thing with members and nonmembers.
DeleteIf a category is innate or obvious from unsupervised learning the giving it a name is trivial.
Neural nets learn categories by abstracting the features that distinguish members from nonmembers and ignoring the rest. This filter generates the CP pop-out effect.
The neural net is not T3-scale, but it can learn (some) categories and does generate CP.
https://www.youtube.com/watch?v=YRD22dY5lck (around the 50 minute mark)
DeleteRegarding the rejection of the Whorf hypothesis that colours are determined by how our culture and language happens to subdivide the spectrum:
I listened to a Radiolab episode a few years ago , in which they talk about a tribe in Namibia that doesn’t have a word for blue, and as such, when shown several blocks of green squares and one blue square (whose difference in size in frequency is not greater than the difference in size in frequency to the other squares), and asked to point to the odd one out, members of that tribe don’t seem to recognize any of the squares as different categorically from the rest.
It also talks about the evolution of colour language – apparently before words for colours enter the language, all languages begin with names for “white” and “black”, and of the colours, red is the first to be named, and blue the last. The hypothesis is that we don’t actually really need names for colours until we can synthetically create them – and because of clay (?), red is the first colour we can make useable in art, and blue is the last.
I suppose this evidence doesn’t contradict what we’ve already discussed: the categories are already prepared (in the case of the Namibian tribe, the borders between categories are not being drawn in places where they aren’t for other people), but it’s that we don’t make use of them until needed.
So regarding Junlin’s comment: how we subdivide and name the colour spectrum, I think does inform our categorical perception of colours, but they are still pre-prepared: it seems that different languages don’t draw the borders in different places; perhaps some borders are left out.
There are some parallel effects in auditory (phoneme) perception. As with color categories (the rainbow), children are born with phoneme boundaries -- e.g., ba/da, ga/ka, ra/la -- but they have a "use it or lose it" property: The boundaries your language does not use are weakened if not lost (as with ra/la in Chinese or Japanese).
DeleteThis might seem like a silly question, but is it possible for the neural nets that generate innate categories to strengthen? I am aware that learned categories require lots of corrective feedback from the consequences of doing the right or wrong thing with members and non-members of a category, but can innate categories become more clear or strengthened by this corrective feedback process? Or another type of process? You mentioned that certain phoneme boundaries have “use it or lose it” properties which can weaken innate categories which would have an effect on your innate CP and this makes sense. But I am curious about strengthening innate CP and to what extent we could do this. Although, even if we are able to strengthen innate CP, I don’t know how useful this could really be for humans.
DeleteI also wonder how neural imaging studies of innate and learned categories can help cognitive scientists. When reading about how certain event-related potentials are correlated with differences in perceived category of the stimulus, I kept thinking back to the reading about mirror neurons. I am inclined to question whether examining the brain and performing neural imagining could help cognitive scientists solve the hard problem. Do these neural imaging studies help neural net simulations?
If neural nets are the right model for category learning, I don't see any reason why they could not strengthen as well as weaken feature weights, whether for new or inborn categories. (New categories, after all, are also "potential" categories within our overall sensorimotor capacities.)
DeleteNeural imaging will not reveal to you the causal mechanism that generates a cognitive capacity (whether mirroring or category-learning), but neural (and behavioral) correlates can help in testing a candidate causal model. That's still the Easy Problem, though. The Hard Problem is a problem for any explanatory model, not just what might be learnt from neural imaging.
“Differences between event-related potentials recorded from the brain have been found to be correlated with differences in the perceived category of the stimulus viewed by the subject. Neural imaging studies have shown that these effects are localized and even lateralized to certain brain regions in subjects who have successfully learned the category, and are absent in subjects who have not (Seger et al. 2000).”
DeleteI have to say that I’ve started my degree being very enthusiastic about what neuroimaging had to offer. In first year courses, neuroimaging is presented to students as a power tool to achieve causation, and as a non-invasive way to empirically test theories about the brain. As for categorization, a big part of what we humans and other organisms do, neuroimaging can only capture a part of the output. Yes, it can tell us that learning has occurred and that this learning is now being reflected by the physiology of the brain. But that’s not cognitive science; that’s not reverse-engineering cognition. Now knowing the power of neural nets to simulate cognitive phenomena, I feel like the time I spent learning neurophysiology should have been spent on computer classes. To go back on what Veronique said, I think it is irrelevant to know whether “the core processes involved in cognizing the in-group and out-group differences in objects/things/concepts/people [are] localized in different regions of the brains”. What we are concerned about is reverse-engineering our capacity to cognate. For that aim, neural nets are good candidate because they at least offer an explanation (that may be right or wrong) while neuroimaging does not offer any.
Some languages have words for feelings or emotions that English lacks.
ReplyDeleteIn English this feeling would be an area of continuous categorization between two concepts where in another language it might be a line for CP of one feeling compared to another. Is this not the Whorf Hypothesis to a degree? It doesn't warp how we see the world to the degree the original hypothesis proposed with color, but it provides extra possible categorization due to difference in language, which alters how the world is seen, as we see the world through these categories.
I'm also curious how the weak Whorf Hypothesis, "language influences our perception of the world" holds up in your eyes. Is what I have laid out above an example of the weak hypothesis? And also how "influence" is a weak word, but you mentioned in this instance, we would come to see "influence" isn't a weak word.
CP induced by category-learning is Weak Whorf. But it requires a measure of similarity within and between categories, before and after learning the category. This would be hard -- Wittgenstein's "private language argument" suggest that it would be impossible -- in case of feeling categories (because of the other-minds problem: how would you detect or correct error?).
DeletePerhaps the distinction between innate and learned categorical perception is a matter of what we are “pre-wired” to categorize (i.e., our capability to categorize) and what we actually categorize. In the example of pitch perception, the basilar membrane encodes the frequency of the incoming sound with limits between 20 to 20,000 Hz. As the basilar membrane is fully equipped by the time we are born, it gives us the capability to categorize. But the learning aspect is what divvies up (specific) ranges of frequencies into pitch classes. Suppose the incoming frequency of a sound is 440 Hz, corresponding to the pitch of A. The fact that we are able to perceive an A and can distinguish it from the perception of a B-flat is innate. However, the fact that we call it an A is learned. Moreover, the fact that we call a sound vibrating at 442 Hz an A is a process of detecting invariances. We assess what pitch category it falls closest to by comparing it to the pitch perceived from 440 Hz. Because the pitch corresponding to 442 Hz sounds more similar to the pitch corresponding to 440 Hz than to the pitch corresponding to 466.16 Hz (B-flat), we categorize it as an A. But suppose we had a different pitch category for every frequency; would that not influence how we learn categories?
ReplyDeleteDiscrimination (relative) and categorization (absolute) are different. It's not clear whether "perfect pitch" is innate or learned (probably a bit of both), but attaching the name to the note is learned, trivially.
DeleteNor is perfect pitch infinite precision. If you have perfect pitch for A440, you can tell if the A is a little high or a little low, but not exactly how much. It's approximate to a semitone, and the rest is just relative to the reference centre semitone pitch. We don't have absolute quarter-tone recognition, just relative.
Professor Harnad, I am quite interested to hear what your opinions are on the following. My research pertains to timbre perception in the Music Perception and Cognition Lab. My supervisor, Stephen McAdams, often tells us that there is no such thing as “the timbre of a clarinet” (or any musical instrument). Rather, each musical instrument encompasses a “universe or galaxy of timbres”. What this means is that we can hear two notes played by the same instrument and perceive them as completely different “categories of timbre” or even coming from two different sources. Suppose we get a clarinetist to play two notes that differ greatly in pitch. We find that, when asked to report whether those two pitches were played by the same instrument or not, people (including musicians) sometimes report that they were played by different instruments. I am curious what your thoughts are on this, given that if each instrument really encompasses a universe of timbres, how can we determine the invariances between any two notes? Wouldn’t we technically need a different “iconic (or echoic) projection” for every note that we hear, even though that seems highly infeasible? Moreover, how can we categorize timbres of musical instruments if it seems that even the within-category differences are separated rather than compressed in this case?
ReplyDelete(P.S. I am sorry that both my comments are music related.)
The question is empirical: we can definitely categorize notes from different instruments correctly (though not perfectly). I don't know how many notes you need to hear for accurate identification. Single notes may not be enough, and extreme pitches and amplitudes my be harder. No doubt the tone-color of instruments (even the piano) change across their range, so maybe reliable identification only works within a smaller range. But from what you describe, that still leaves more than just a "universe of timbres."
DeleteThe test should not be "do these notes come from the same instrument?" but "name the instrument plaing these motes."
Neural net would be interesting here: Come to the lab and test our neural net for how well it can learn to identify different instruments, and how.
(No need to apologize for musical examples! We use them in the lab too, as in Grace Wang's experiment.)
Though it’s certain that there’s infinite variation in the sounds any instrument is capable of producing, I feel like referring to the timbre of an instrument as a distinct, observable category probably does more good than harm. A composer orchestrating a piece will employ such categories; the sound they hear in their head --how a melody will sound on clarinet(s), cello(s), any other instrument(s)-- benefits from their conception of how different instruments sound when played conventionally. I agree with prof. Harnad’s point that single notes might not be enough for recognition. I’d bet that, hearing a short phrase (maybe even a weird leaping one without repetition or conventional voice leading), most people could tell if it was played on one instrument or if it were distributed across a few. Maybe people are good at identifying timbre in music because melodies mostly move in smaller, singable steps; if you were to contrive a melody of mostly giant leaps across an instruments range, people would probably have a much harder time finding a common thread in timbre.
DeleteI’d be really interested to see if a neural net might shed some light on what features might be used to discriminate between instruments. I feel that I’m pretty powerless to describe the timbre of instruments that I listen to or even play, especially If I weren’t allowed to narrow in on properties like attack & decay, which can’t be the whole story. Also sort of related: I wonder if such a neural net test could turn the question of what makes a ‘nice’ instrument empirical - if you played guitars ranging from broken beaters to top of the line instruments, would there be some general observable trend in the features of the sound they produce?
Some of these questions could perhaps get an approximate answer today from neural nets, but of course neural nets -- especially for sensory input -- are only as good as (1) their sensors (peripherals) and (2) their input representation (coding). Neither of these is as yet very advanced today. So what they tell us about timbre might be more about the limitations and assumptions of (1) and (2) today than about what the ear and brain really does. But worth trying anyway. A potential McGill Honours thesis, worth doing with Steve McAdams at CIRMMT (where Erica is doing her doctorate), and maybe jointly with me.
Delete“It will have to be shown that our perception of the world can also be warped, not just by how things are named but by what we are told about them.”
ReplyDeleteI can't help but relate this to an article I read for another class (Cohn: https://www.jstor.org/stable/3174209). I don't know if you are familiar with that, but she basically compares natural language and nuclear analysts' language, which is incredibly specialized and narrow compared to the former. She gives many examples, one of them being how warfare creates death and horror in English, which has emotional valence, whereas in the highly technical language of the professionals, it only creates "collateral damage", which is undesirable, yet emotionally neutral. She spent a few months among those men and talks about how the language used to describe the events changes the way civilian death feels (or doesn’t). I wonder if that would be a good example (though anecdotical) of the effect of “what we are told about things” insofar as if it were the case that we were told that collateral damage and death both elicit a sense of sadness or whatnot, this thing would have felt the same instead of differently?
Re: “I wonder if that would be a good example (though anecdotal) of the effect of “what we are told about things” insofar as if it were the case that we were told that collateral damage and death both elicit a sense of sadness or whatnot, this thing would have felt the same instead of differently?”
DeleteThis is super interesting comment! I am not sure it would be an example of the effect of “what we are told about the things” though, I believe it would still remain a vocabulary effect. The phrase “civilian death” includes the concept of man, someone like me, trying to go about their life. On the other hand, the term “collateral damage” also encompasses damage made to buildings and objects. This means that as a result of compression/separation effects, “civilian death” feels closer to my own death, whereas “collateral damage” feels more like the destruction of property. To me, we would still be in the case of our perception being impacted by how things are named, rather than what we are told about them.
I also had to read this article for class, and I think it may be more about euphemisms than categorical difference. Euphemisms 'broaden the scope' of a phrases 'meaning' so to speak, in order to offer the listener more possibilities to what something means. In other words, it is a less precise way of saying something in order to indirectly get at the emotional content. 'Death' in warfare could be described as 'collateral damage' (although grossly inhumane and poorly worded). But since 'collateral damage' could refer to any number of contexts, and thus, meanings (warfare but also gambling, for example) it is not necessarily another category of thing, but a *broader* category of thing.
DeleteThis brings us back to the idea of Frege's sense and reference: 'death' and 'collateral damage,' in the context of military warfare have different senses but the same reference. Another example would be 'the queen of England' and 'Queen Elizabeth II' -- different senses, but the same reference.
I don't think the article necessarily has to do 'with what we are told about things' in regards to categories, but rather the ability to manipulate language to serve abstract purpose (which is the basis of poetic, non-kid-sibly language). This however is still an incredibly powerful skill of language, it evokes 'feeling' beyond 'what it feels like'.
Interesting questions, but not, I think, a way to approach the question of the effects of descriptions (propositions) versus names on perception -- and especially not on the perception of categories of feeling, which is a very special case (because it is subjective and internal rather than objective and measurable).
DeleteThere is no doubt that what we say (language, including connotations and associations) can bias perception, in various ways, though, and it has been studied in the context of categorical perception and social stereotyping by Tajfel and those that have since followed up his work.
Categorical perception is the phenomenon when awareness of one category changes to awareness of a different category instantaneously, instead of occurring gradually. According to the article, this phenomenon occurs “whenever perceived within-category differences are compressed and/or between-category differences are separated, relative to some baseline of comparison.” I’m having trouble differentiating these two instances of categorical perception; can someone give me an example of each (i.e. one example of compression and one example of separation).
ReplyDeleteHi Taylor,
DeleteThis is a really great question. I am unsure as to how accurate my answer will be but I hope that Professor Harnad and/or our fellow classmates will correct me or add on if I misinterpreted anything. From what I understood, within-category compression and between-category separation describe the processes by which we assimilate or differentiate instances into their respective categories. Professor Harnad wrote another article in which he re-words the two types as "when members of the same category come to look more similar" (within-category compression) and "when members of different categories come to look more dissimilar" (between-category separation).
Using colour perception as our example, perhaps we can think about the grouping of distinct shades of red as "red" as within-category compression; we are "compressing" the unique characteristics of each shade within a single category. I believe that Funes, who perceived "every instant [as] infinitely unique, would not be able to generalize in this way. Then, between-category separation would be illustrated by distinguishing the cut-off point between red and orange. At what point does red stop being "red" and should be named "orange"? Or alternatively, when does "pa" become "ba"? This process emphasizes the boundaries or differences between categories, whereas within-category compression focuses on the space within a category. I hope this helps!
Taylor, the reply of Vivian is correct: categorical perception (CP) is not a sudden but a gradual change as you learn (some) new categories. Be careful not to confuse categorization with CP.
DeleteExample of innate CP: the rainbow: Greens look more like greens (compression) than blues (separation), even when the wave-length difference between them is the same.
Example of learned CP: Perez-Gay, F., Thériault, C., Gregory, M., Sabri, H., Harnad, S., & Rivas, D. (2017). How and why does category learning cause categorical perception? International Journal of Comparative Psychology, 30.
(By the way, the way, Maddy Gregory, Hisham Sabri and Dan Rivas are all alumni of this course, and worked under Fernanda Perez-Gay, who got her PhD in IPN last year.)
“In this respect, the "weaker" CP effect for vowels, whose motor production is continuous rather than categorical, but whose perception is by this criterion categorical, is every bit as much of a CP effect as the ba/pa and ba/da effects.”
ReplyDeleteI have noticed that when a Mandarin learner is trying to mimic a sound that does not exist in her native language, she can do a better imitation if she “does not overthink it”, especially for vowels (which is within expectation). She can repeat it quite well if she “just repeats it” and does not put the vowel into her native language’s vowel space.
This seems to suggests that, to some degree, she is able to perceive the difference between the foreign sound and the (most) similar one in her native language and even produce it, but by putting it into a category that is different from the one a native speaker of Mandarin has for the same sound, the critical differences are “ignored” and lost.
I have a similar experience when I was learning English. For me, “not treating the sound that I am imitating as a sound in a language” gives me more accuracy in the imitation. It is as if I am deliberately breaking down the boundaries and put it in a continuous spectrum, and re-construct what the categories are.
It's vague what the strategy is to "not put the vowel into her native language’s vowel space," but it may have something to do with Shklovsky's notion of ostranenie, "defamiliarization" or "making strange" in art, by portraying or perceiving things out of their usual context. It is very likely that learned CP is context-dependent, because things may need to be categorized differently in different contexts: An apple when you're looking for something to eat or you're looking for something to hold down a piece of paper in the wind (to give a example of short-term contexts, but you could find long-term examples too, such as the vowel in "turn" or "girl" which a native anglophone cannot extract from the retroflex vocoid "r" in order to produce the short first vowel in (German) "Götterdämerung" or the long one in "Goethe."
Delete(That said, CP effects were always relative differences in degree between vs. within categories, rather than all-or-none, like a cliff! There are shades of /ba/ and /ga/ just as there are shades of blue and green. They're just perceived as smaller than when they cross the category boundary. Ventriloquists can probably also do a good approximation in shades of production ba/ga production, despite the "motor theory of speech perception," though of course not as good as with truly continuous vowels.)
This comment has been removed by the author.
ReplyDeleteWhorf is not about whether language changes behaviour but about whether it changes their perception of the world. If religious teaching (words) made people hallucinate things that were not there, that would be a Whorfian effect.
DeleteI don't think that would be considered a Whorfian effect. The Whorfian effect posits that language alters our perception of the world, for instance, having different names for the colors blue and green is what allows us to perceive these as categorically distinct, however, other societies who do have have different words to distinguish blue from green wouldn't be able to categorize the colors blue and green and would therefore not have a categorical perception that distinguishes a clear boundary between blue and green, that is, they would see those colors as the same, or at least as belonging to the same category. Thus, two people speaking different languages would come to see the world differently from one another. However, this has not been supported in research.
DeleteIn the case of people with schizophrenia, language isn't really a factor in their hallucinations, and their religious beliefs may play a part in shaping the content of their hallucinations, but this would not be considered a Whorfian effect. I think the problem is that Nathan proposed that religious views change how people perceive the world similarly to how Whorf believed that language changes how people perceived the world, however, Nathan is using the word perception differently than Whorf was. Nathan is using the word perception in the sense of our views, opinions, and beliefs about the world, whereas when Whorf used the word perception, he meant quite literally our perception in the neuroscientific sense, that is, how we categorize our sensory input of the world.
In this paper, Harnad mentions this about within-category compression and between-category separation:
ReplyDelete“This evolved into the contemporary definition of CP, which is no longer peculiar to speech or dependent on the motor theory: CP occurs whenever perceived within-category differences are compressed and/or between-category differences are separated, relative to some baseline of comparison. The baseline might be the actual size of the physical differences involved, or, in the case of learned CP, it might be the perceived similarity or discriminability within and between categories before the categories were learned, compared to after.”
This is an interesting theory of categorization and I am just curious that if it forms a contrast or is consistent with Wittgenstein's theory of family resemblance for categorization. Wittgenstein uses the example of “game”, from which he argues that no one can pick out the essential common feature of that classifies a ‘game’, but there are a series of overlapping features that none of which necessarily apply to every instance of games.
I wonder how would the contemporary CP theory and Wittgenstein’s theory possibly relate or contrast to each other.
“I wonder how would the contemporary CP theory and Wittgenstein’s theory possibly relate or contrast to each other.”
DeleteI think that this comment is interesting in the sense that it asks the question about vague categories like the one of “games” (even though if its members are quite concrete). The process of “features abstraction” could be more flexible than a rigid set of features; maybe, let’s say, you need some “core features” + a certain amount of peripherical features that can vary from one member to the other, and the “compression” would be the mean to eliminate the non-relevant peripheral features. But maybe the specific example of “game” provided by Wittgenstein is not the best because games are highly conventional and arbitrary (we make rules **so as** we can play). So maybe the CP is not relevant here. Maybe it is just a mere vocabulary effect of naming certain activities “game”.
But I think it is still interesting because it asks the question of the nature of the features that are abstracted and used in cognition in general.
What Wittgenstein allegedly showed was that the "classical theory" of categories -- that categories have features that are necessary and sufficient to being members -- was wrong.
DeleteHis example -- games -- was meant to show that there is no feature that all games share, so categorization is based only on "family resemblances": some members have some features, others have others, but they're still all in the same "family."
Wittgenstein was talking about metaphysics: "What kind of a thing is a category?"
Cogsci is not about metaphysics; it's about the mechanisms that underlie organisms' capacity to categorize ("do the right thing with the right kind of thing"). As such, there has to be a right and a wrong, otherwise there's no category.
If people can sort things (correctly) into categories -- e.g. games and non-games -- then there must be a set of features (possibly a complex Boolean combination of them: and, or, not, if/then) that will make it possible to sort them all correctly. (If only mostly correctly, then some of the boolean rule could also be statistical rather than exact.) "Family resemblances" is just an either/or rule "apples are red or green"): no use looking for a color they all share (as Jerry Fodor did with his "vanishing intersections" argument). The feature that all members share is that they all conform to the Boolean rule.
In cogsci, there is nothing wrong with the "classical theory." See Appendix 1 of the categorization skywriting in 6a.
“CP occurs whenever perceived within-category differences are compressed and/or between-category differences are separated, relative to some baseline of comparison” (Harnard 2003)
ReplyDeleteThe questions that I had posed in my 6a sky writing this week are mostly answered in the following reading (Harnard 2003) for 6b. Here we can see that the nature of Categorical Perception is that of a discernment device that works along the boundaries and similarities of things; either emphasizing sameness, or delineating points of difference. Furthermore;
“There are even recent demonstrations that although the primary color and speech categories are probably inborn, their boundaries can be modified or even lost as result of learning, and weaker secondary boundaries can be generated by learning alone. (Roberson et al. 2000)” (Harnard 2003)
Consequentially, the mechanism does lend itself to efficient categorization, and exhibits an innate flexibility with regard to the integration of knew knowledge schemes through learning. Resultantly, biases can be enhanced and muted through learning, and are already present within evolved categorical perceptions.
Categorical perception is an indirect effect of changing the weights of features.
DeleteThe Whorf Hypothesis, in the context of color perception, argues that how we subdivide the color spectrum is arbitrary given that it is defined and a learned result of our cultural surroundings. While I disagree with the hypothesis in this case (such as with Berlin and Kay), I do believe that our environment and culture can vary how and what we categorize; e.g., with language. Although Harnard makes clear that language-induced CP-effects remain to be clearly demonstrated (and I cannot be sure if language by itself can help us acquire categories), I do feel that language can still play a role in varying how we cognize (and thus, categorize) the sensory inputs that we are perceiving. For example, the French language is based upon grammatical gender. Due to this, whether a noun is characterized as either masculine or feminine influences the form of other words in a sentence related to it. Related to this, Melancon & Shi (2015) showed that French children processed better both familiar and newly learned nouns if they were gender-matched. Due to this, they concluded that grammatical gender is represented abstractly across all words within the French language and is actively used when perceiving novel words and combinations. Other languages, such as, English or Chinese, do not possess this grammatical gender. I wonder whether these syntactical differences also differentially modulate how we perceive and ground sensory inputs in our surroundings (e.g., people, places, things, time). If so, what is occurring in the case of bilingual individuals that speak both a language that possess and does not possess grammatical gender?
ReplyDeleteDoes grammatical gender influence the way we see the world? Sun and moon are masculine and feminine in French and the reverse in German, and neither in English or Chinese. How does this affect perception?
DeleteAs a student of a French immersion school from age 4 to 18, I can attest to some of the difficulties students faced associating gender in French when their English equivalents had none. On a societal level, there are some words that have a somewhat patriarchal assignment of gender given to them reflecting their "lesser than" status in the world or that the object was better suited to a woman than a man for whatever silly reason.
DeleteI would argue that the grammatical gender *does* influence the way we see the world and creates an unconscious bias in the way that things are perceived. Since this is a learned concept, not only does it bias whatever object is being considered, but it may also reflect on the expectations of gender overall. Similar to how meaning is grounded in other meanings, if an object is grounded in a gender, it does help to inform that gender going forward.
Perhaps to better illuminate this issue we could create a distinction between seeing and feeling. By this I mean that perhaps grammatical gender influences the way we *feel* about the world but it does not influence the way we *see* the world. It is obvious that language has an effect on how we feel and can determine how we feel. Ascribing gender to every object would impact how we feel about or ascribe certain properties to that gender. I agree with you that grammatical gender does influence the way we *feel* about the world. However, does this mean that the way we see (in the perceptual way) is changed by the fact that we could assign masculine or feminine traits to a certain object? Does our ideas about masculinity and femininity affect the way or ability for us to *see* the sun? I would argue that although language may change and influence how we feel it may not have the ability to let us see more or less things based on the vastness or compactness of our language.
DeleteFlora, I find your distinction between feeling and seeing interesting. However, I think categorization can not only change how we feel about the things we categorize, but also our sensory perception of them. After learning to categorize seemingly indistinguishable shapes into two distinct categories, we will perceive the shapes that belong to the same category as more similar to one another, while the members of different categories are perceived as more different from each other. This is a direct consequence of our category-learning device weighing certain features, for example those that distinguish the members of the two categories, more heavily than before learning took place. I understand that for male and female noun categorization this seems less evident (and may not necessarily be the case), possibly because of the rather arbitrary nature of how their members and non-members are categorized.
DeleteHi Kyle, Flora and Erica,
DeleteI also find the distinction between feeling and seeing interesting. However, both categorization and categorical perception seems originate from how we do what we do, (ie the easy problem). I cannot think how the categorization could change the feeling, without touching on the how we do what we do. Even if it is innate or learned, categorization capabilities seem to be corresponding to the things we do, not what we feel. And saying that something affects the feelings (without altering the doing mechanism), even though we have no idea about causality of feeling, might not be appropriate.
Clearly though, as Kyle indicated above, people do feel some effect caused by the language: for instance, the associating gender in French. However, I am hesitant to say that this effect is solely due to Categorical Perception that is driven by how things are named. It looks our ideas about these words are not solely influenced by categorical perception but most probably it is caused by what we are told about them.
I think there is not inherently bad about associating genders with names. There wouldn’t have been any bias due to associating some words more with a woman than a man (or vice versa), if our perception of the genders were not already warped by what we are told about them (Not just from how things are named). I think that it is more than just being categorical perception that is driven by how things are named. So, I think maybe the full-blown language effect may explain such an effect of language better.
Apologies for the delay on this skywriting. It will be addressing the second paper “How and Why Does Category Learning Cause Categorical Perception?”
ReplyDeleteOne of the first things that comes to mind when hearing the ““right thing to do” is whatever the organism needs to do with the members of the category and not-do with the non-members of the category” (Perez-Gay et al., 2017), ie., the catch phrase of categorization is Kluver Bucy Syndrome. This is a neurological disorder where patients have difficulty recognizing objects, and therefore have the tendency to put everything in their mouths or by displaying inappropriate sexual behaviours. This demonstrates how important categorization is because it shows the result of what happens when it is not done properly. From what I understand, patients are mis categorizing because they are improperly detecting the features of a “thing” and therefore cannot distinguish members from non-members of a category. Patients also definitely do not experience within category compression and between category separation.
A question I have about the paper is with regards to the neural net model. Is it not possible for the neural net model to be used in the reverse engineering problem because of the fact that unlike the neural net model, human data shows that there is either no correlation or a negatively proportional relationship between difficulty and size of categorical perception effect?
Agnosias and anomias are perhaps more relevant to categorization disorders.
DeleteNeural nets can learn to detect features and thus categorize. The relation between difficulty and the size of the CP effect is still under study.
Re: Categorical Perception
ReplyDeleteI understand that CP could be formed not only by experience, but by some innate characteristics of the human sensory system as well. I think language goes a step further than simply helping us create categories. Language allows us to recognize which referents may be innate versus which ones are learned. For example, everyone can agree on what blue is, what is hot, or what is cold. In contradiction, cross cultural notions of beauty, or wealth may vary and have different referents. Are there any which are purely one or the other, or is CP always a combination of innate structures and learned differentiation criteria?
Pauline, if I understand correctly categorical perception could be completely innate, learned or a combination. For instance, categorizing faces (or perceiving all the elements that make a face as a face) or distinguishing blue-green seems innate to humans while, like you point out, categories based on beauty and wealth are perhaps purely learned. But even within innate categories such blue-green, the degree of blueness between a cyan or teal seems to be learned. To me most categories (and hence categorical perception) seemed to have been purely learnt. And like you, I find the role language plays in CP fascinating. Although the Whorf effect seems to be limited, Harnad seems to be indicating that there still could be a full-blown language effect. I wonder what the current status of this research is.
DeleteWhen categorical perception happens, the perceived differences within members of a category become smaller and the perceived differences between members of different categories becomes larger. Some categories, like color, are innate and some are learned. At the end of the article, Harnad discusses whether or not categories (and CP) can be acquired through language alone, and not direct sensorimotor experience of the learner.
ReplyDeleteI really do not see how we would not be able to acquire categories through just instruction.
On a different note regarding this passage, " Some things are more big and some things are less big. In this case the category is "continuous" (or rather, degree of membership corresponds to some point along a continuum). There are range or context effects as well: elephants are relatively big in the context of animals, relatively small in the context of bodies in general, if we include planets."
We discussed in class that categorization is distinct and absolute, and that continuous things are not categorical. So I'm confused by this excerpt, how a category is continuous if the previous sentence is true.
Hi Anna,
DeleteMaybe I can be of help here (of course, maybe not).
Categories like “big” and “small” are relative assessments. When we say “Elephants are big” we are not saying “Elephants belong to the set of things that are big”. We are saying “Elephants are big compared to the things I have in mind, which are smaller than elephants”. Thus, the distiction “big” is not of a categorical nature, but a relative one.
Even if we were to consider “things that are smaller than an elephant” to be a category, we would learn very little from this. There are infinitely many things that are smaller than an elephant, and to relative degrees: a human; my Flintstones gummy vitamins; that specific hydrogen atom over there. (Another thought: Categorizing is learning to do the right thing with the right kind of thing. The designation “things that are smaller than an elephant” does not entail a “right thing” to do with it.)
I disagree with you, Paul, when you say, “The designation “things that are smaller than an elephant” does not entail a “right thing” to do with it.” In what way is “things that are smaller than an elephant” different than the category “apple”? It all depends on the context in which you’re using it. “Things smaller than an elephant” is an ad hoc category, as you are creating a category that could be of use in some situation (doing the “right thing” with this category). Categorizing these things requires the same sort of feature abstraction as categorizing apples (with different features, of course).
DeleteHowever, I am similarly confused, as the existence of continuous categories like "big" and "small" directly contradicts what Prof Harnad said in class about categories being distinct and absolute. The only thing that I can think of as a solution to this contradiction is to say that “big” is categorically distinct from “small”, but that within these categories there is a continuum of degree of membership.
I'm not sure if the proposed solution works, as things that are not in the category "big" are not automatically "small".
DeleteI think in class Prof Harnad was discussing the "categorical" variety of categories ("all-or-none" membership). He distinguished this phenomena with "continuous", but maybe left out that continuous categories can still be seen under the larger umbrella of categories.
"There are still the Lane/Lawrence demonstrations, lately replicated and extended by Goldstone (1994), that CP can be induced by learning alone."
ReplyDeleteIf CP can be induced by learning that means that our brains already have this underlying capability to do so. If that's the case, then perhaps studies like these aren't showing us that CP can be learned, but rather that it can be activated from a dormant state. I think these studies are evidence of innate/evolved CP rather than learned CP.
"And there are also the countless categories catalogued in our dictionaries that could not possibly be inborn."
If most of the words in our dictionary are just variations of a basic set of words, then those words are being categorized into different variations of those basic words. So wouldn't these categories be innate rather than acquired? I personally don't think they are innate, I'm just confused because this thought makes me think otherwise.
I read in an article that every year we create around 5400 new words, and about 1000 of those stick. If someone is able to know the meaning of one of these new words without ever having heard it before, then what does this mean? I think if a large percentage of individuals could actually do this then there would be quite significant evidence of something, I'm just not sure what.
In regards to your first reply, just because something can be learned does not mean the learned process was previously dormant in our brains. We can learn how to do calculus, but the ability to do so is not sitting idly from birth. Maybe smaller parts of a learned phenomena are supported by innate abilities, but the ability itself is not simply waiting for activation.
DeleteI had a similar confusion in conceptualizing testing learned CP, as it is easy to think "if participants are already innately demonstrate CP, wouldn't that just come into play when testing for learned CP?" However, it helped me to think about how each of these things (learned vs. innate CP) are tested. Vowel sound CP was shown in infants, rendering it innate. Whereas if you look at Harnad’s description of the typical learned CP experiment, you will see that it entails learning new categories and assessing if the “within-category-supression or between-category-separation” takes place. The existence of innate CP for certain categories (i.e. speech and color) does not automatically extend to all existing categories, most you must learn.
I think what I was getting at, but my thinking wasn't quite there yet, is that what's dormant is the ability of us being able to do CP, not the function itself being dormant. So I agree with you that the entire thing isn't just embedded in our heads.
DeleteLearning about CP made me realize an experience I had when I was in high school. I had learned what the word petrichor was and that it meant the smell after it rained. I was aware of the fresh smell beforehand, but I didn’t have a name or a category to put it in. I knew it smelled earthy but other than that, I didn’t have a category. However, once I learned what petrichor was, I believe that I experienced the weak Whorph. I was better able to pick out the smell of petrichor and I better differentiate what was the smell of petrichor and what smell was just earthy. Harnad, in your reply to Gabriel and bringing up the private language argument by Wittgenstein, I hope that I understood what the argument is saying. Is this stating that although there is a commonality in language, each person link between the word and referent are different and therefore we each possess a private language that we effectively use to translate the common language that we hear. In this case, it becomes difficult to know whether our categories of feeling are universal. We may all feel but we don’t know if that feeling is different or the same?
ReplyDeleteJust to be sure: Categories are the groups of similar kinds of things, in which membership can be either "all or none", or along a continuum. This is different from both the original and contemporary definitions of categorical perception, which is a sensorimotor effect. Originally, CP referred specifically to the stark distinction between the qualitative features of one category versus another, and this was understood to be an all-or-none effect. Now, CP describes the phenomenon of differences within categories being less noticeable +/- differences between categories appearing more obvious. And this compression/distinction process is evident after having learned the categories? Except for colors and speech sounds, because we have an innate ability to perceive these according to categories.
ReplyDeleteI am not completely certain I understand what is being described as “neural nets”. Is this a computational model/simulation or a description of brain mechanisms? I believe it’s the former, but I’m asking for clarification. And what does the “I/O” signify in “neural net I/O models”?
I am also unclear on why the Whorf Hypothesis is rehabilitated. Is it because categories and categorical perception could theoretically be learned via language?
In general, categorization is all-or-none, not relative or a matter of degree. Similarity is relative.
DeleteCategorization (and category learning) are not the same thing as categorical perception. Sometimes, when a difficult category is learned, members of different categories look more different after you have learned the category, compared to before.
The categorization is all-or-none. The between-category and within-category similarities (relative) change (probably because your category-learning device has detected and more heavily weighted the category-distinguishing features).
(The mistake -- or exaggeration -- about CP [in phoneme perception -- but not in color percetion] had been that some people thought people could only tell the difference between members of different categories, but not between members of the same category. The truth turned out to be a matter of degree: degree of pairwise similarity. But categorization itself is all-or-none and absolute, not relative. A fish is a fish and a bird is a bird. Not that some things are more fishy and some things are more birdy...)
Neural nets are computer models that learn, and can be interpreted as a nework of interconnected units whose connections can be strengthened or weakened. They learn to detect and filter features. No evidence for them in the brain yet.
I/O is Input/Output (as in "weak equivalence").
Strong Whorf/Sapir is wrong: Naming things differently does not make them look more different. But learning to categorize things (do the right thing, including naming) can sometimes change perceived between- and within-category similarity (because of feature detection, weighting and filtering). That learned CP is a Weak Whorf/Sapir effect.
"We have seen pictures of them, but what had those who first drew those pictures seen? And what about categories I cannot draw or see (or taste or touch): What about the most abstract categories, such as goodness and truth?"
ReplyDeleteHow do the way we categorize and use language affects our perception of what is true from false, wrong from right? CP is learned so what about fundamental concepts about justice, injustice or values or what is possible from not possible? Where is the line drawn ? A lot of it is tighten with the fact that we learn those categories from what other people have told us about the world, but then, when we can't experience these concepts directly, how can we make up our own mind without heavily being influenced by societies norms views and conditioning? I think that this is one of the main problems of categorization and a very interesting one to study across cultures and time. For example in the previous DSMs being a part of the LGBTQ community was categorized as a part of mental disorders, thus members of this category were stigmatized being mentally ill and it was "wrong", and today they are not categorized that way anymore so we perceive members of this community differently and not as wrong anymore. Categorization has so many societal implications.
Hi Andrea, I agree with you that categorization can have high societal implications. For instance, research shows that the legalization of same-sex marriage in the US was associated with reduced anti-gay bias (https://www.pnas.org/content/116/18/8846/tab-article-info). The change in categorization of something from illegal to legal had huge social repercussions. And the example you provide of DSM categorization of LGBTQ proves the same point. In these scenarios, the categories people form seems to be highly influenced by societal rights and norms. Simultaneously, it is affected by each individual's sensorimotor experiences. And that variability is perhaps a result of a long chain of past sensorimotor experiences of concepts/members within or related to a category.
DeleteHi Anna, I also agree with the impact of categorization on society. I feel like the reason why simply naming things differently does not make them more different (strong Whorf/Sapir Hypothesis) is accurate and it relates to computation in my opinion because giving a name to something when this name is not encoded and learned, it has no value. For instance, when you mentioned that the legalization of same-sex marriage in the US was associated with reduced anti-gay bias proves the strong Whorf/Sapir hypothesis. When something because legalized we don’t simply give a name to it, it becomes a law and people use and learn from laws to guide their behavior in order to be a good citizen.
DeleteIn this article Harnad succinctly explains the basis of categorisation and categorical perception. I enjoyed learning the evidence for categorical perception based on varying theories, computational models and neural data. I understand that when a neural net learns categorisation and shows categorical perception (within-category compression and between-category separation), the weights/bias towards certain features (and hence nodes or connections) increase in comparison to others. Is there a way we can correlate this with changes in neural connectivity in the brain when we learn? Although it won’t be a one-to-one mapping, based on the studies that report changes in ERPs that correlate with differences in the perceived category, are there studies correlating the neural network biases with ERP changes? Although neural nets are mathematical models, I’m trying to understand if they can bear resemblance to neural connections in the brain.
ReplyDeleteIt is also interesting that categorical perception leads to reduced cognitive effort. It struck me as an example of one of the ways “lazy evolution” gets away with not coding all categories in the human DNA. And by not coding it can allow individuals to learn flexible and new categories depending on the requirements of their environment. And perhaps the enhanced flexibility does lead to a higher chance of survival. In this regard, perhaps evolution is being more efficient overall.
“to show that it is a full-blown language effect, and not merely a vocabulary effect, it will have to be shown that our perception of the world can also be warped, not just by how things are named but by what we are told about them.” (Harnad, 2003b)
ReplyDeleteThe talk by Lera Boroditsky showcases how language is able to change the way we think of the world. However, her examples are all related to additional/different vocabulary certain languages afford. The Austrialian Aboriginal language Guugu Yimithirr use cardinal marks for navigation – they do not have words for left and right. Yet, this doesn’t truly warp their perception of the world; the Guugu Yimithirr have a learned sensorimotor categorical perception. To think is to cognize and categorization is the major component of cognition. Boroditsky misses the mark: languages do not induce different categorization – vocabulary does. I still don’t know what it would be like to find proof language induces different categorial perception. Does this lead us back to the Other-Minds Problem?