6b. Harnad, S. (2003b) Categorical Perception
Reading: Harnad, S. (2003b) Categorical Perception. Encyclopedia of Cognitive Science. Nature Publishing Group. Macmillan.
Instructions for commenting: Quote the passage on which you are commenting (use italics, indent). Comments can also be on the comments of others. Make sure you first edit your comment in another text processor, because if you do it directly in the blogger window you may lose it and have to write it all over again.
“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.”
ReplyDeleteThis passage nicely shows what categorical perception is: the same physical world can be sensed either continuously (smooth, gradual spectrum with no clear breaks) or discretely (distinct, separate categories with sharp boundaries) depending on our perception. Differences between items in different categories are exaggerated in our perception, while differences within the same category are compressed. I love how this quotation outlines a perfect example of this: continuous perception (many shades of gray) compared to categorical perception (distinct red, green, and blue). Even though colour is physically continuous, categorical perception shows how the boundary lines of categories can shift depending on perception, highlighting how our sense of similarity among the “right kinds of things” can be warped to make categories feel more real in our heads. This also plays a role in symbol grounding, since categories must be “grounded” (to abstract the features of members from no members/sensorimotor learning) in perceptual distinctions, and categorical perception helps ‘structure’ these categorical members and non-members in a way that supports our unique perceptions as individuals living in the world.
Rachel, yes, categorical perception, whether inborn, through evolution, or learned through trial, error and feedback, makes categories "pop out" in our perception by selectively enhancing features that distinguish categories from one another and suppressing features that do not distinguish them. (It's a little like selective attention, when someone says "sort the marbles quickly by their color: ignore differences in size.")
DeleteRachel, I really like the way you explained categorical perception with the shades of gray vs. color example. The example worked so well with Harnad's (2003b) notion that the same physical input is experienced either continuously or as different types of categorical experiences, depending on how our perception organizes it. Your example beautifully illustrates how our minds exaggerate the differences across category types and compress the differences between categories making sharp distinctions sharper than they are in the physical world. This connects well to how categorical perception structures the categories that we later use for language and symbol grounding.
DeleteProfessor, your analogy of selective attention helped me to think of categorical perception as more of an active filtering process. The reading stresses that categorical perception , acts by enhancing features of a category that distinguishes a category or category type and downplays features of a category, that do not make a distinction whether the categorical distinction is innate to the perceiver, via categorizing types that were learned through feedback. That filtering makes categories "pop out" perceptually, thereby allowing us to quickly sort the world and respond. It is not just a perceptual act; it is also a key cognitive mechanism of categorizing observable input, thereby influencing how we ambitiously carve up reality and attach words to those sub-carved objects.
Lorena, another excellent synthesis. Important to grasp that whether the category-distinguishing features that make the categories "pop out" like colours can be "learned" by evolution (how?) or learned in an individual's lifetime (how?).
DeleteAnd the potential categories (how many are there of them?) are rarely regions of a spatial continuum, like wave-length, separated by a threshold: The potential features that divide potential categories are often discrete categories themselves too (like colors, shapes, or textures). (Category membership, after all, is all-or-none (i.e., categorical, like "bird," or "feather"; not a matter degree, like "big.")
A good way to set your intuitions is to open a dictionary (or encyclopedia) and see how many words are likely to have innate, evolved feature-detectors, and how many are more likely to be learned in a lifetime by exposure (unsupervised learning) and trial and error, with feedback from consequences of doing the right and wrong thing (as with the mushrooms on the island).
Thank you, Professor! I appreciate that you described that differences across the categories are either a part of evolution or are picked up through our experiences in life. I can understand how some features that define categories, such as colors or faces, may have been built in because they served an adaptive purpose. Other features, such as figuring out which mushrooms are edible, seem to be more about figuring things out from trial and error along with feedback throughout our lifetime. Your point about most categories not just being points on a simple spectrum also stood out to me. A lot of them are based off a discrete features such as, “has feathers” or “does not,” that I think also fits with the all-or-none category membership. To think about categories and concepts in this manner (discrete transition) helps clarify how categorical perception is not simply the sensory sorting of input, but actually how our evolution and our experiences come together to create the way in which we split the world into categories.
DeleteLorena, and perhaps the most interesting thing is the way direct sensorimotor learning capacity, mirror-capacities, interaction and communication with kin and kind (in social species) led to the capacity for language (so far only in humans),
DeleteThe capacity to perceive, detect, abstract and learn category-distinguishing features directly evolved hundreds of millions of years ago. The capacity to learn categories indirectly from verbal description of their distinguishing features (by someone else who knows them) is thought to have evolved only a few hundred thousand years ago and again, only in our own species in the Great Ape Family.
(Please don’t misinterpret these affirmations about our species as admiration…)
Thanks for your feedback! I find the connection you made interesting. I'm intrigued by how you linked the evolution of direct sensorimotor learning to the eventual evolution of language. It makes perfect sense that the ability to form categories based on direct experience (like learning what to do with what) had to evolve in the human brain before we got to the point of sharing, in spoken words, that we could decorate the world with verbal descriptions. The fact that verbal grounding which is learning categories purely through someone else's description, evolved much later heavily implies that language only added another layer on top of earlier mirror capabilities and social learning.
DeleteIt is fascinating to think the capacity to communicate about categories, instead of only observe or act on categories, was the defining thing that allowed knowledge to flow socially and culturally. The difference seems like saying " I can learn what a fruit is by tasting it.” and "I can describe to you what a fruit is, so that you don't have to taste it." That transition from single-source learning to shared-output learning, seems to me to be the tipping point of human cognition.
***EVERYBODY PLEASE NOTE: I REDUCED THE MINIMUM NUMBER OF SKYWRITINGS. BUT THE READINGS ARE **ALL** RELEVANT TO AN OVERALL UNDERSTANDING OF THE COURSE. SO, EVEN IF YOU DO NOT DO A SKYWRITING ON ALL OF THEM, AT LEAST FEED EACH READING YOU DO NOT READ TO CHATGPT AND ASK IT FOR A SUMMARY, SO YOU KNOW WHAT THE READING SAID — OTHERWISE YOU WILL NOT HAVE A COMPLETE GRASP OF THE COURSE TO INTEGRATE AND INTERCONNECT FOR THE FINAL EXAM.***
DeleteAs I read this article for some reason Frank Jackson’s “Mary’s Room” thought experiment sprang to mind. I think it can be used nicely in support of the paper’s overall argument. As a recap: Mary is a neuroscientist who knows everything about the science of color (she’s read all about wavelengths, neural processing, etc) but she has lived in a black-and-white room for her entire life. One day, she experiences red for the first time. In seeing red, she learns something new: the experience of what it is like to see red. This showcases the gap between facts/data vs. having direct interaction/perceptual experience. But we can take it further: only now in seeing red does she know the perceptually-grounded categorical differences between red, black, and white. This is important for her understanding of red. Without CP, Mary could have had all the knowledge for distinguishing red from black, but not have been able to truly know where the perceptual boundary for where “this is red” ends and “this is black” begins. AI models trained only on symbols lack the sensorimotor experiences of the categories themselves (no grounding of the concepts). They only “know” like Mary does before seeing red. Although AI might be able to simulate categorization, without grounded CP, it, like Mary, cannot truly have an understanding of the categories...
ReplyDeleteThat’s a really good point, Elle!
DeleteThe article on Categorical Perception explains that perception isn’t continuous, our brains divide what we sense into distinct categories with sharp boundaries. Before Mary sees red, she only knows about color through facts, but she can’t actually experience where one color ends and another begins. Once she sees red, she forms that categorical boundary separating red from black or white. This is what Professor Harnad means when he says real understanding comes from grounded perception, from direct sensorimotor experience rather than information alone. Like you said, AI or Mary-before-red can process data, but without perceptual grounding, they don’t genuinely understand the categories they use. It also makes me wonder, if we show ChatGPT an image of something red, does that truly count as perceptual grounding, or would genuine grounding require the kind of active, sensory interaction with the world that allows meaning to emerge from experience rather than mere data?
Elle, good attempt at a synthesis, but you are mixing up "categorical perception" (CP) with perception itself (i.e. feeling: the fact that it feels like something to see (or hear or taste), whether in black and white or technicolor or stereo).
DeleteDoes T3 grounding have to be felt -- or is sensorimotor feature-detection and action (doing), zombily, enough? If you have a clear answer, you've solved the HP...
Ayla, same questions for you: T3 sensorimotor grounding might be necessary for understanding what content-words and propositions about the neurophysiology of color vision refer to and mean, but does the T3 grounding need to be felt? Why?
Think about it; it will wean you off the neuroscientist-Mary koans...
As Elle and Ayla have both mentioned, the paper shows that categorical perception (CP) is about how our minds divide the continuous world in discrete kinds, helping us do the right thing with the right kind of thing. But I think the key is that this is not the same as what it FEELS like to perceive those kinds. The professor describes CP as “within-category compression and between-category separation”, which can be innate (ex. speech, color) or learned through experience.
DeleteI am not familiar with Frank Jackson’s “Mary’s Room”, but from what has been said, it sounds like Mary’s case is about the feeling of perception, not CP itself. Even before she “feels” red, she could still know “where “this is red” ends and “this is black” begins.” I think what this paper suggests is that CP can exist functionally, as a way of sorting the world, even without consciousness.
On the point about T3, I don’t think T3 grounding has to be felt. A T3 system could have CP in the sense that it can warp continuous input with compressions/separation yet still be a “zombie” with no feeling. So, Professor, am I understanding correctly that CP is about how perception is organized rather than what it feels like, and that Mary’s Room is about the latter, which is a different topic?
Annabelle, good question. Despite its widespread usage, I don’t think the “colorblind neuroscientist Mary” thought-experiment gives much insight. It just conflates “knowledge” and “know-how”.
DeleteT3 robotic grounding provides know-how: what to do with what: how to detect the features that distinguish the right ones to eat and from the ones not to eat. Grounding also gives the robot the T2 capacity to call them “edible,” and to describe to someone who does not yet have the know-how, what are the features that distinguish the “edibles” from the “edibles” — all the way to being able to give a university course in mycology.
But that does not mean the T3 robot (professor) has felt what it feels like to see, or touch or taste or smell the features of edible and edible mushrooms (or to feel what it feels like to perceive any feature, of anything).
So without the WW “know” — which conflates T3 know-how with felt know — the Mary thought experiment is just saying the “knowing” is not just “know-how” if it is not felt.
But the Hard Problem of explaining how and why anything at all is felt is left untouched.
Know-how (i.e., doing capacity) is the province of the “Easy Problem.”
(BTW, you could already say the same thing much more simply if you imagine a color-blind lipstick-designer who can only see black, white and grey — but really does see them, because he does feel what it feels like to perceive the differences among those shades. He could design (chemically) formulas that sell really well, although he himself cannot see the difference.
Same point as the Mary experiment: Feeling is not know-how. The same thing applies to the Mycology Professor, if he is just a grounded T3 zombie! And “knowing” is a WW.)
Thank you for that distinction Professor! it helps clarify the difference between categorical perception and perception itself. Categorical perception is about how we learn to carve the world into meaningful divisions. On the other hand, perception, in the deeper sense, is the felt experience of those divisions.
DeleteIf a T3 system can perceive sensorimotor regularities and act on them, it could in theory learn to distinguish “red” from “black” by detecting patterns and responding appropriately. But that kind of grounding, though functional, wouldn’t necessarily involve internal feeling. It would be a form of understanding based on performance, not experience.
So the question changes and becomes “does understanding require that internal, felt dimension, or is successful doing enough?”
My sense is that T3 robotic grounding can explain how meaning connects to the world, but without feeling, it still lacks what makes understanding conscious. In other words, T3 grounding might solve the symbol grounding problem, but not the hard problem of what it’s like to perceive.
In the article, the categorical perception of abstract concepts like “goodness and truth” are questioned. This is quite interesting as it prompted me to think about how humans would ground such non-concrete ideas. How does one see or experience goodness? And further, there isn’t a universal definition of goodness. We need to break down the concept of goodness into a series of boolean combinations in order to ground an abstract concept. What would this proto-goodness grounded experience be? I found this article very well structured and now that I understand the concept of categorical perception better, I only have more questions!
ReplyDeleteI like your point, Jesse. Harnad really does leave us wondering how abstract ideas like “goodness” or “truth” could ever be grounded. He argues that even those categories must, at least indirectly, be rooted in sensorimotor experience. Through language, we construct” Boolean combinations” of simpler grounded categories; so, for instance, good might inherit its meaning from experiences of comfort, safety, or approval, mediated somehow through words. Thus, categorical perception for abstract ideas isn’t about seeing them directly, but about how language reshapes perception by linking new meanings to old, embodied ones.
DeleteJesse C and Gabe, please have another look at the earlier replies on “abstract categories”.
DeleteA dictionary will never betray you (just as the mushroom island won’t).
I think you understand now what direct sensorimotor grounding (DSM) refers to: It’s mushroom-grounding, where you learn by trial-and-error sampling to detect what are the features that distinguish the edible ones from the poisonous ones.
The Mycology professor is like the dictionary (or encyclopedia, or manual, or ChatGPT): they can define the distinguishing features — on condition that you already know the referent of the content-words in the definition. Otherwise you have to look them up too, and so on. That is indirect verbal grounding
(IVM).
In DSG you learn to detect and abstract the distinguishing features directly. In IVG you are told them verbally.
Once someone has directly grounded enough content-words (a “Minimal Grounding Set”), all the rest can (in principle, but not necessarily in practice) be done purely verbally — combining and recombining words that refer to distinguishing features.
All categories are “abstract” (because they are all based on selectively picking out the category-distinguishing features, and ignoring the rest). (Did you read Funes the Memorious? Week 6a)
I was thinking about what you said about “goodness” being hard to ground. Maybe what Harnad means is that even if we can’t see or touch “goodness,” our brains still use experiences to make sense of it. Like, when something feels right or kind, we learn that it fits the “good” category. So maybe abstract ideas are just made from smaller, simpler experiences that we already understand. It’s clear we use language to build on smaller ideas to create bigger ones, like truth or justice, even if we can’t directly sense them.
DeleteHarnad’s section about “resolving the “blooming, buzzing confusion” touched on how our brain’s supposedly have the innate function to detect faces. As psychology students, I’m sure the majority of us have been taught about the fusiform face area. This area in the temporal lobe is responsible for recognizing and processing faces. This is why we so easily recognize or “make” faces out of things that are not human or animal, like seeing a face from the windshield and head lights of your car or the windows and front door of a house. There are other areas in the brain that seem to be specific in categorical recognition, such as for locations and written words. These are broad categories however, and in arguing if things categories are innate or learned, I think it may be a mix that relies on direct experiences and perceptions with said categories.
ReplyDeleteKaelyn, as usual, a dictionary will not let you down:
DeleteSample a dictionary to estimate the proportion of content-words whose referent is likely to have an innate feature-detector like the fusiform gyrus. Look at the definitions that are unlikely to have innate feature-detectors in our brain. What you will see is definitions composed of multiple content-words, for whose referent it is unlikely that we have evolved direct, innate, composite feature-detector.
But as you keep tracing their definitions, you will find that you reach words that are more likely to be directly groundable. And many will also be the names of sensorimotor features, likes colours, shapes, tastes, textures, sounds, for which we do have evolved, selective feature-detectors “learned” (by evolution) across millions of years. We may not yet have coined a content-word for them, but they are ready to be lexicalized that way by giving them a name.
Harnad’s article shows how categorical perception allows us to sort the world efficiently, exaggerating differences between categories and compressing those within. But I think there’s a hidden trade-off here: what subtle distinctions do we lose when our minds sharpen boundaries? By focusing only on differences that matter for categorization, we risk overlooking important within-category variation. For example, the way artists notice shades others miss, or how rigid speech categories can mask emotional tones.
ReplyDeleteThis makes me wonder if true perceptual intelligence isn’t just about drawing crisp boundaries but knowing when to soften them. Categorical perception is powerful, but adaptive minds might be those that can tune their boundaries for the task at hand, allowing for both efficiency and richness of experience
Alexa/b>, many learned categories are attentional, based on selective feature detection and weighting. We have a category for dogs, which distinguishes them from cats; but we can alsohave categories for different strains of dogs, and even sub-strains. So the features that are ignored when we are trying to distinguish dogs from cats may have different weights when we are trying to distinguish Whippets from Weimraners.
DeleteThe space of potentially learnable categories is not continuous (like the wave-lengths of light) but consists of (approximately) discrete features (or “dimensions”) like shape (straight/curved, concave/convex, pointed/blunt, moving/still), which are themselves learnable, nameable categories.
So learned CP is much weaker and more modifiable than biologically evolved CP: The between-category-separation/within-category-compression effect arises from “feature-reduction” (or dimension-reduction:what is that?), from learning to selectively enhance some features and ignore others, in some contexts and not others.
6.b. Harnad explains how categorical perception helps us tell things apart by exaggerating differences between categories and minimizing differences within them. It’s a neat idea that links perception to language and learning, but it also feels a bit simplified. Real-world perception seems messier—boundaries aren’t always clear, and context can change how we see things. The theory works well for sounds or colors, but it’s less convincing for complex categories. Still, it’s a useful way to show how our brains shape what we perceive, not just what’s “out there.”
ReplyDeleteIt’s important to recognize that CP isn’t just defined as a fancy “trick” the brain can accomplish but instead it reflects how perceptual systems are shaped by both innate mechanisms and experiences. Although it is true that boundaries aren’t always entirely clear, they can sometimes be fuzzy, CP still illustrates a crucial principle. Even when boundaries are messy, CP shows how our brains take differences between categories, exaggerate them, and minimize differences within them, helping us make sense of continuous variation. Even in these ambiguous situations it shows that our brains don’t simply reflect the external world but actively structures sensory input crucial for learning and language regardless of complexity.
DeleteEmily, good reply to Anthony. And for the complex (and increasingly abstract) recombinations of features that can be expressed and transmitted by words (indirect grounding) the possibilities become enormous.
Delete
ReplyDeleteBy starting from Liberman’s all-or-none view, the concept of categorical perception became much clearer for me. Harnad's article explains that Liberman believed, based on the ba vs. pa syllable experiments, that individuals are unable to discriminate items within the same category, which led him to adopt the all-or-none view on categorical perception. However, Harnad explains that categorical perception is a process in which within-category differences are compressed rather than extinguished and between-category differences are separated relative to a point of comparison. Building on this, I have come to realize how important context and goals are in influencing what features I pay attention to when interacting with things. For instance, at a grocery store, I might distinguish Macintoshes from Royal Galas based on their slight differences. On the other hand, when I make an apple pie at home, I might use any apple, compressing within-apple differences, because my goal is different.
“Some of our categories must originate from another source than direct sensorimotor experience, and here we return to language and the Whorf Hypothesis…”
ReplyDeleteHarnad’s example of the frog’s inborn ability to detect “flies” illustrates evolutionary category preparation—nature’s way of pre-wiring perceptual biases that ensure survival. Such innate detectors show that categorical perception long precedes language or culture, grounding the ability in action- and survival-relevant perceptual distinctions, from an evolutionary perspective. Still, as Harnad suggests, it seems that “some of our categories must originate from another source than direct sensorimotor experience,” with language being the prime suspect. If language bolsters our ability to perceive and understand categories, and strong categorical perception contributes to survival and evolutionary success, then it makes me wonder whether this ability (i.e. language-induced CP) is just a consequence/by-product of language or actually contributed as a driver of the evolution of language in humans.
Liam, you raise an important point about the difference between categories we learn through direct experience and those we gain through language.Prof. Harnad would agree that some of our categories (like “fly” for a frog) are rooted in sensorimotor experience and evolutionary preparation. These categories come from direct interaction and survival feedback. But this alone can’t explain the vast number of categories we have, like justice or bachelor. Those come from language.
DeleteLanguage lets us learn new categories through what Prof. Harnad calls symbolic theft—we can pick up knowledge from others’ words instead of discovering everything ourselves. Once some words are linked to real experiences, we can combine them to create new, more abstract ideas. These language-based categories even start to shape how we see and understand the world, a process Harnad calls language-induced CP.
To your question, I believe that Prof. Harnad would say this ability wasn’t just a by-product of language but a key driver of its evolution. Language induced CP gave humans a huge adaptive advantage, allowing us to share and use categories instantly rather than relearning them through experience. Still, as Harnad insists, “it can’t be theft all the way down”: every word, no matter how abstract, must ultimately trace back to something grounded in the sensorimotor world.
Gabriel, good grasp and observations. Look also at the power of deep learning models.
DeleteLiam, the capacity for category learning (through feature detection and abstraction) evolved in many species (hence it began before language in ours). Category learning sometimes (not always) involved perceptual or attentional changes (learned CP effects). Words can certainly induce attentional changes: It’s still an open question whether words alone can induce perceptual changes (as in hypnosis, which is closely connected to language).
Lucy J, good points, but language’s main power seems to be in transmitting categories (from those who already know the category’s relevant distinguishing features to those who do not). That’s not yet creating categories; but of course there’s category discovery and creation too, in science and art and invention.
But let’s not forget the destructive power of language, especially today…)
What is Baldwinian Evolution?
This comment has been removed by the author.
DeleteA Baldwinian evolution refers to the evolving of a specific trait or ability that is not initially inherited but rather learned and honed over generations, ultimately providing the species with an adaptive edge. A fake, but fun example is the following: Recently, people have come up with a theory suggesting that crocodiles can now mimic a drowning person (by putting a limb out of the water) to lure humans in. This has been debunked by experts, but if it were true, then this tactic would certainly be the result of baldwinian evolution. That behaviour is not genetically encoded like a newborn's crawl to its mother's teats. It would have likely randomly yielded positive results (lured humans in), providing crocodiles that can do it an adaptive edge, and over generations, their offsprings would have adopted and perfected the tactic, making it second nature.
DeleteThis article touches on categorical perception in both what we can see (colour gradients), what we can hear (ba/pa continuum) and the distinctions that arises from the synthesis of all our sense that intake stimuli to categorize the words we can find in a dictionary. With all of these sense in play, and how each stream of input goes through wildly different biological mechanisms to be perceived by humans, it makes me wonder if one sensory modality is better at performing this compression and separation than another, would that make the categorization more granular and is granularity itself something that is positive in terms of our cognition when we are talking about categorization?
ReplyDeleteLucy M, What is categorization, category learning and categorical perception? Kid-sib could not grasp what you were asking here.
Deletecategorization, now that it’s been explained through the lectures and skywritings, is learning how to do the right thing with the right kind of things (edible (to eat) -> apples, oranges or writing utensils (to write with) -> pens and pencils). Category learning, understood through machine learning models is that they are trained to understand the “right thing” and the “right kind of thing” mentioned above, this comes both through direct information (being told that fuzzy, four-legged thing that barks is a dog) or through an gradual inference (a lot of the same shape all close together is ‘something’ – the name of that something is not important, that comes after, just knowing those are something together is what is important here). Finally, categorical perception is how our brains can properly categorize, without the exaggeration and eliminations of differences a lot of what we perceive will exist on a gradient without clear boundaries (like colours) and the nice discrete buckets we’d like to throw everything we perceive into will no longer exist.
DeleteCategorical perception is an important consideration in designing a T3 system, which aims to model how humans learn and represent categories through both sensorimotor grounding and language. CP shows how experience organizes perception into meaningful categories, reflecting the T3 principle that cognition arises from interaction and learning rather than pure computation. By grounding knowledge in sensory and motor experience, the T3 system models how higher-level understanding can emerge without needing to replicate the brain’s physical structure.
ReplyDeleteSierra, how are categories, category learning and CP related to T3 and grounding?
DeleteThe idea of Learned CP shows how experience can reshape perception, but if language itself has some innate foundations, it might act as the bridge between what’s natural and what’s trained. Our brains could be prepared to categorize sounds or meanings in certain ways, but language experience fine-tunes those boundaries over time. In that sense, innate biases give us the framework, while learned CP fills in the details. Together, they show how perception and language co-evolve. They’re partly hardwired, but are continuously refined through learning and use.
ReplyDeleteSannah, the capacity to learn language is innate, but language itself is learned. Some categories (like colors, facial expressions and phonemes) are innate or partly innate, but most are learned. They can be learned directly, through sensorimotor trial and error and feedback from the consequences; or learned indirectly, by word of mouth from someone who has already learned the category (directly or indirectly). The direct and indirect learning is learning to detect the category-distinguishing features. (Please make sure you understand that.)
DeleteI think that also shows how language doesn’t just describe categories but actually shapes how we notice them. If we only learned through direct experience, we’d be limited to what we’ve personally seen or done. But with language, we can form categories we’ve never experienced ourselves. So learned CP isn’t just about fine-tuning—it expands what we’re even able to perceive
DeleteCategorical perception is described as a way our minds carve the continuous world into discrete categories, allowing us to respond appropriately to different kinds of things. What stood out to me is that CP doesn’t just make perception more “organized”, it also makes cognition more efficient. By compressing within-category differences and exaggerating between-category ones, CP effectively reduces the cognitive load where we no longer need to process every tiny variation, only whether something crosses the categorical boundary. This links well with symbol grounding: grounded categories allow a system (biological or robotic) to act correctly without infinite perceptual fine-tuning. It also raises a question that could CP evolve primarily as an efficiency mechanism, rather than as a means for “understanding”? If so, even a completely unconscious T3 zombie could have CP, because it’s a functional adaptation, not a phenomenological one.
ReplyDeleteRachel Y, all good points. Some details:
DeleteCategorizing (=?) is provisional and approximate, not exhaustive and absolute. (It reduces uncertainty but not to zero, which is only possible for (1) the formal, syntactic propositions of mathematics: those are necessarily true on pain of contradiction; and for (2) the Cogito (Sent) (=?)).
Learning category-distinguishing features is a form of "dimensional reduction" (what is that?) (Funes the Memories).
(a) Category learning, (b) categorization, (c) content-word grounding and (d) even categorical perception (between-category separation and within-category compression) -- are not necessarily felt (sentient) rather than just done. (If you could explain how and why they are felt you would have solved the "Hard Problem" (Week 10)).
Harnad's article introduces categorical perception (CP) as a qualitative form of categorization, in which individuals perceive sensory information – whether that be speech sounds or visual stimuli – in discrete categories. This allows us to detect larger differences between categories (e.g., “ba” versus “pa”) and smaller differences within categories (e.g., two slightly different versions of “ba”). A key takeaway is the underlying nature versus nurture debate – that is, whether CP is an innate biological adaption or learned through experience.
ReplyDeleteThis makes me question how perceptual conditions like synesthesia – the ability to merge multiple senses, such as seeing colours when hearing sounds – fits into the mold. Unlike CP, which sharpens perceptual boundaries, synesthesia blurs senses together. Indeed, strong genetic components are at play, but said pairings are only formed through exposure and experience. Therefore: If synaesthesia reflects an evolved trait, why doesn’t everyone experience it? Or, if it is something that is rather learned through association, why doesn’t everyone acquire it?
Interesting point of how synesthesia seems to go against the goal of categorical perception. While it's meant to act like a mental filter, synesthesia merges different senses like sound and color thus blurring the lines. I think you are right to question why we don't all have synesthesia, and I think the answer lies in an evolutionary trade-off; since thinking is fundamentally categorization, then the most crucial survival skill is abstraction. Our brain needs to selectively ignore or forget the chaos within every moment in order to detect reliable patterns. If one fails to filter out these chaotic details, then you run into problems such as lack of ability to generalize.
DeleteGeneralization is crucial for efficient categorization in order to be able to identify which action to take given the situation. Which I believe suggests that synesthetisa is a byproduct of complex brain wiring, not a primary feature that species were designed to use
Grace, CP can be either inborn or learned. Most of our categories are learned rather than inborn. Synesthesia does not blur one sense with another; it is just an accompaniment.
DeleteZafina, read "Funes the Memorious". Learning, abstraction and generalization require the selective enhancement of relevant features the selective ignoring of irrelevant. There is a trade-off between exact memory for detailed features and selective memory for the relevant features (to be able to categorize: "do the right thing with the right kind of thing"). Ask GPT about "hypermnesia."
What stood out to me is how categorical perception connects to how we build knowledge itself. If our brains naturally group things into categories, that means even our first experiences are already shaped by some kind of sorting system. We’re not just seeing the world as it is, but through the filters we’ve learned or inherited. That makes me think learning isn’t just adding new information, it’s changing how we divide things up in our minds. So when language adds new categories, it’s almost like it gives us new ways to see reality, not just new words for it.
ReplyDeleteRena, What is categorization? What is category learning? What are the two ways to learn categories? Language does not create categories (except in creative writing): It transmits them from those who have learned (or created) them to those who have not.
DeleteCategorization is doing the right thing with the right kind of thing. Categorization can be categorical (all-or-non), in which membership is absolute, such as the example of the bird, where something is either a bird or not, or it can be continuous, which means it varies by degree, such as the size of something that can be either big or small. However, many sensorimotor categories are a mixture of the two.
DeleteCategory learning is the process of categorization, where we learn how to sort things into categories and distinguish one category from another other which is also known as categorical perception. There are multiple ways to learn categories, such as evolved CP, which is built in our sensory systems, and learned CP, which is developed through our sensorimotor learning, which includes language-induced learning.
“Categorical perception occurs whenever perceived within-category differences are compressed and/or between-category differences are separated.” Its so cool that our brains “warp” perception to make the world more discrete and manageable. It makes me wonder how much of what we see as distinct categories (like colors or sounds) is shaped by the structure of our sensory systems versus by language and learning. If categorical perception can be both innate and learned, how much of our perception is biologically fixed, and how much can culture or experience reshape it? Could we, for ex, learn to perceive new categories that don’t currently exist in our sensory world or are we ultimately constrained by evolution?
ReplyDeleteShireen, I really like how you framed categorical perception as the brain “warping” reality to make it more manageable, that’s such a vivid way to put it. Your question about how much of perception is fixed versus shaped by experience made me think of a study we read from Kuhl, showing that even animals like chinchillas experience categorical perception for speech sounds. That suggests some categories are truly built in. But I also think language and culture can stretch these boundaries, for example, after learning a new language, we learn to hear subtle differences in sounds that don’t exist in our native language (like the difference between the "u" and "ou" sound in French). So maybe evolution gives us the starting categories, but experience keeps reshaping them.
DeleteShireen, What is categorization? What is category learning? Categories can be learned by an individual brain during a lifetime (edible mushrooms), or through millions of years of genetic evolution (colors).
DeleteCognitive science is not ontology: It does not tell you what does or does not "exist" in the "world". Cogsci is trying to reverse-engineer how organisms can learn what to do with what kind (category) of thing (including what to call it). And for that, you have to be able to learn to distinguish what is and is not in the category. If a "category" does not "exist," how can you tell what's in it?
Jad, both the perception and the production of categories (in mirror domains like speech) are based on features, whether sensory or motor features (or sensorimotor "affordances" that depend on both). And an organism can learn to detect a category's distinguishing features c during a single lifetime or feature-detectors can evolve genetically in a population across across generations.
Look in a dictionary or an encyclopedia and count what proportion of nouns, verbs, adjectives and adverbs refer to a category whose distinguishing features are known at birth, and what proportion need to be learned through trial/error/feedback (as on the mushroom island).
"Liberman et al. (1957) reported that when people listen to sounds that vary along the voicing continuum, they hear only ba's and pa 's, nothing in between"
ReplyDeleteI found this passage really interesting because it reminded me of a video I saw in one of my other psychology classes about the McGurk effect, where someone’s lips say “fa” but the sound says “ba,” and our brain ends up hearing something in between. That video showed us how perception isn’t just about the sound entering our ears, but also about how it is shaped by our brain. In this passage, even though the change from "ba" to "pa" is gradual, our mind draws a sharp line and decides if the sound is a "ba" or a "pa" with nothing in between. It’s like perception snaps things into place so we can tell them apart. We said in class that anything that falls on a continuum is not categorization, but here our brain creates the categories itself. I think this reveals how our minds simplify continuous reality into clear parts and I find that very interesting.
In reality our perception doesn't mirror reality, reality is more of a “continuum of blooming, buzzing confusion” but what we see are categories caused by CP. The fact that some of these categories are inborn like ability to recognize faces while some others are explained by the Whorf hypothesis and shaped by language makes me wonder to what degree are these categories flexible. If our brains are wired to perceive categorically and we have certain categories that we have been utilizing to our whole lives can we train ourselves to see the world differently? Our perceptions and probably change gradually over time without us thinking about it and from new experiences and new acquired knowledge but to what degree can we control our categories? Could we make ourselves experience the world in a different way? Maybe what I am talking about would be similar to cognitive behavioral therapy…
ReplyDeleteWhorf believes that colors are perceived categorically because we have NAMED them categorically. This appears to be wrong since even across cultures we categorize colors similarly which indicates we all perceive color relatively similarly regardless of language. However to me, this is very reminiscent of an experiment where infants ( Korean or english speaking) were categorizing containment as either loose fitting or tight-fitting but English adults did not while korean adults still did (in korean there is something in the language that differentiates between tight and loose containment). Basically, CP may not only be induced by language since we all have the necessary sensorimotor tools to perceive any categories. Yet, language may make us more sensitive to some categories over others. We can all see the same colors and we can all see the tightness of containment, but language could make the difference of salmon and tea rose or a book in bag and a book snug in your shelf more readily available for CP.
ReplyDelete(if anyone is interested this is the link to the experiment I mentioned:https://www.sciencedirect.com/science/article/pii/S0010028502005145)
Building on Alexa’s point about how sharpening conceptual boundaries can cost nuance, I can’t help but think of the tension between two models of cognition: one that survives by compressing distinctions and another that evolves (learns) through revision. CP works well when categories stay stable, but in practice, learning happens instance by instance, through a lot of information and overlapping cues. If the features that define a kind of thing shift, as they often do in real environments, then compression can stabilize too early. Learning categories, as the mushroom island example Harnad presents, doesn’t happen in a vacuum or an isolated state of attention. As for mushrooms or berries– texture, season, and co-occurrence doesn’t draw a single boundary (ie. ‘red berry’ =bad; ‘blue berry’ =good) but a moving one, co-indexed with everything else around it.
ReplyDeleteThat seems to sit uneasily with T3 grounding. If T3 depends on sensorimotor feedback, it needs the efficiency of CP and the ability to unlearn it. Should a grounded system detect when its distinguishing features have stopped working, and update them without starting from zero? Does CP, by definition, trap it in its own past successes? My fear is that this is a question for the hard problem—because the awareness required to reassess and adapt a category, implies feeling: it feels like something to be wrong.
In the reading, it was suggested that some categories don’t come from sensorimotor experience but rather language. For instance, we have never seen or touched categories like truth and goodness, but we can understand and engage with them meaningfully. It was said that neural network models show that once some words are “grounded” in sensory experience, they can be combined linguistically into more abstract concepts, therefore, can language not only determine how we describe the world around us, but also perceive it? And further, if language can generate new categories and even warp perception, could it also be responsible for creating entirely novel modes of thought?
ReplyDelete« Some of our categories must originate from another source than direct sensorimotor experience, and here we return to language and the Whorf Hypothesis: Can categories, and their accompanying CP, be acquired through language alone? »
ReplyDeleteThe passage cited above discusses and wonders whether CP and categories can be acquired through language alone. The ability to form higher-order categories remedies the problem of categorizing things we have not grounded in a sensorimotor manner such as a unicorn, but it does not answer whether we wouldn’t know what a unicorn was if the word “unicorn” was not in our languages’ dictionary. If I take an example off of another language, the word “komorebi” in Japanese refers to the light leaking through tree leaves. Even though we do not have a word in English to name that phenomenon, it is still possible to categorize it or “name” it as light leaking through trees. We still know what it is and can abstract the distinctive qualities of it without having a specific and distinctive name categorizing it. Moreover, to acquire CP and categories directly through language, a minimal amount of content words need to be grounded to be able to further ground categories through heresay. So, if starting with absolutely no grounded categories, it does not seem possible from what we know to be able to acquire CP and categories solely through language, but is seems possible to categorize something without it having a distinctive name to designate it, a name that would be found in a dictionary for example.
This paper offers a rehabilitated Whorf Hypothesis: language may not determine what we do or do not perceive, but it can amplify or blur distinctions by modifying categorical perception. This made me reflect on how language might shape the way we experience our emotions. If a language has only one word for “happy,” then a broad range of positive feelings might be perceptually compressed and expressed through that single label. Conversely, if a language has many different words for “happy,” each capturing a slightly distinct nuance, its speakers could more finely categorize and thus more precisely feel their emotions.
ReplyDeleteIt seems intuitive that speakers of such a language would have greater emotional depth, perhaps even richer inner lives. For instance, Spanish reportedly has nine positive words for every negative one, while Mandarin’s ratio is closer to three to one. If the rehabilitated Whorf Hypothesis is correct, could this imply that Spanish speakers, whose language allows finer distinctions among positive emotions, might actually experience happiness more diversely - or even more often?
Much of categorical perception is due to the compression and separation of perceived differences of items in categories. So, items in the same category will seem alike, as we compress their differences, and items in different categories will seem distinct as we separate their differences. For example, if we have three colour swatches: a true red, a red with some amount of yellow mixed in, and another red with twice as much yellow mixed in, which we then perceive as orange. We will not see a great difference in the yellowness of the two reds, but the orange will seem much more yellow than either of the reds, even though they physically have the same difference in yellowness.
ReplyDeleteWhat I found interesting about Professor Harnad’s paper was how perceptual “warping” can apply to abstract categories that we can’t directly interact with, such as a unicorn or goodness/truth. The paper suggests this might be due to Language-Induced CP. This is a process where basic content-words, which have referents and are grounded, are acquired through sensorimotor experience, are then combined into complex combinations.
ReplyDeleteThe paper gives the example of the category "bachelor," which is defined by combining the categories unmarried and man. Computational models suggest that these higher order categories inherit the CP effects of compression and separation from their simpler components and some categories can even add "a layer of separation/compression of its own".
This reading talks about how we manage a world that is really overwhelming. Our minds use something called categorical perception (CP) which treats differences abruptly and qualitatively (like colours) rather than continuously, (grey shades). Categorical perception is the brain's way of warping reality where we end up compressing within category differences and/or separating between-category differences. Despite this generalized function of compressing and separating perceived differences, CP was initially explained exclusively by the Motor Theory of speech perception. However, this theory was abandoned because CP was found in infants before speech and even in animals like chinchillas, showing just how it is not unique to human speech and may be an innate, evolved preparation. Now, given that language-induced CP effects for abstract categories (like 'goodness' or 'truth') are proposed to be inherited through the linguistic combination of grounded sensorimotor categories, how can empirical studies definitively demonstrate the existence and influence of CP for these highly abstract categories when the effects themselves still "remain to be directly demonstrated in human subjects" and might be inseparable from the inherited sensorimotor biases?
ReplyDeleteAs we are now more advanced in the course content, I'll allow myself to try to link this paper to Week 9 – Chomsky. I am wondering if the idea of “innate categorical perception,” paired with the example of language, can be related to or used as an argument for Universal Grammar, which is also said to be innate. The paper notes that infants exhibit CP for speech before they begin speaking, suggesting an inborn perceptual structure. Could this imply that UG's innateness can partially be explained through categorical perception shaping how infants parse linguistic input from the start?
ReplyDelete