Harnad, S. (2005) To Cognize is to Categorize: Cognition is Categorization, in Lefebvre, C. and Cohen, H., Eds. Handbook of Categorization. Elsevier.
We organisms are sensorimotor systems. The things in the world come in contact with our sensory surfaces, and we interact with them based on what that sensorimotor contact “affords”. All of our categories consist in ways we behave differently toward different kinds of things -- things we do or don’t eat, mate-with, or flee-from, or the things that we describe, through our language, as prime numbers, affordances, absolute discriminables, or truths. That is all that cognition is for, and about.
We organisms are sensorimotor systems. The things in the world come in contact with our sensory surfaces, and we interact with them based on what that sensorimotor contact “affords”. All of our categories consist in ways we behave differently toward different kinds of things -- things we do or don’t eat, mate-with, or flee-from, or the things that we describe, through our language, as prime numbers, affordances, absolute discriminables, or truths. That is all that cognition is for, and about.
After reading this article, I find myself agreeing with the notion that cognition is mostly (if not only) categorization.
ReplyDeleteRE: "But the categorization problem is not determining what kinds of things there are, but how it is that sensorimotor systems like ourselves manage to detect those kinds that they can and do detect: how they manage to respond differentially to them."
How do we discriminate from the light input on our retina which we call a person and another light input which we call an appliance? With my understanding of this reading and previous neuroscience courses, I’ll take a stab at this.
Humans are wired to categorize. The more that we see similar patterns (i.e four limbs, a face, body parts in specific proportions) the more that we will categorize it into one stimuli (“one chunk”). As adults, when you see someone you typically see the entire person and not just a combination of body parts. However, the first time you see a chair you might see it as a combination of shapes and it is not until that combination becomes familiar enough to you that you will “see” the entire object. In this example, categorization is learned. Once we correctly assume that something is a chair and react to it appropriately, we get a reinforcement signal that strengthens the synapses that were active, thereby reinforcing our assumption. The next time we see the leg of a chair, it will activate one of the synapses which will in turn activate the other ones that it has previously fired with and we will be able to recognize the leg as a part of a chair. The more that we detect the same reoccurrence the quicker we will be to categorize it as "one chunk" and not a combination of wood. The times we are mistaken in our categorizations, we are punished, dopamine is withheld and the synapses are weakened. We learn categories per what is salient to us (i.e. we can interpret micro expressions on someone’s faces) but we would not be able to differentiate between a male and a female pigeon unless the reduction of uncertainty was salient or served an evolutionary function.
Elise, it's not just from seeing things over and over (unsupervised learning) that we become able to categorize them. A big part comes from the corrective feedback from the consequences of categorizing correctly or incorrectly (supervised learning). The most powerful current model of this is "deep learning," which is a combination of unsupervised and supervised learning (and has one of its big hubs here in Montreal, in the lab of Yoshua Bengio). Deep learning (which is usually simulated computationally, rather than being a real parallel, distributed multi-layered neural network) is a potential candidate for the mechanism that finds the features that distinguish the members from the non-members of a category. But a lot of the T3 hardware is also in the dynamics of the robot's sensory and motor systems.
DeleteFrom what I understand, unsupervised learning is foundational for supervised learning. For corrective feedback to impact learning systems, we must first be able to recognise kinds and thus, categorise it for them to be right and wrong. At the same time, right and wrong are abstract categories that are grounded in unsupervised sensorimotor categories. Then it seems to follow that we must necessarily acquire the categories of right and wrong categorisation without supervision and so, supervised learning is not necessary for categorisation, although it does make categorisation undeniably more efficient.
DeleteNot related to the reading, but for those interested - Yoshua Bengio will actually be a keynote speaker at this year's NiRC (National Integrative Research Conference) on March 9th!
DeleteYi Yang, you may be right that unsupervised learning (passive exposure) needs to precede supervised learning (trial and error with corrective feedback), but how can you ground category names (what's and X and what's not an X) without corrective feedback?
DeleteThanks Liza. Bengio's keynote is March 9 2017.
Fascinating, Stevan, this is just a thought. I'm not sure if any of what I said is correct but I was just entertaining the thought that: if we gave our subject unsupervised learning just by passively exposing them to all ranges of our stimuli, let's say playing randomly each of the audio stimuli we generated with the same multidimension setting. It seems that they will be able to perhaps map out each stimuli on a continuous spectrum and describing as the sounds on this end sounds more flat and muted than the sounds at the opposite end, which sounds brighter.
DeleteThey may not have a categorical name to each end, but they will seem to have a dichotomy and an idea of where each sound lies. Does producing this dichotomy through training in unsupervised learning count as a "weak" categorization? I know the spectrum idea is continuous and relative, so it really violates the all-or-none rule, such that maybe we can't call this categorization at all?
In reconsidering what I just wrote, maybe that dichotomy and the mapping of the sounds at the radical end of the spectrum is perhaps just a product of our built in categorical perception to properties of sounds, which is not at all an evidence that the subjects did learn with merely unsupervised training...
My main problem with this article is that while it gives a clear account of what cognition is and does, it still seems to fall in the same trap as behaviorist theories: it merely give descriptions and no explanation. Cognition is categorization, fine. But we get back to the same question: how do we categorize? I just feel like we are simply switching concepts and observing mechanisms under different angles without looking inside these mechanisms. Computationalism at least has an explanatory ambition, but I don't see in what way categorization theories provide more information than behaviorism.
ReplyDeleteMael, have a look at "deep learning," as above, if you want a candidate mechanism. Nothing behavioristic about that...
DeleteWhile reading this paper I had the same confusion as Mael. Without a definition of cognition, it is very difficult to understand the claim "cognition = categorization".
DeleteIf we find that other species categorize, is it safe to conclude that they cognize. For example, birds categorize. When a bird builds a nest it separates nest building material from non-nest building material. Crickets categorize sound frequencies. Crickets create an abrupt boundary in the frequency continuum around 16 Hz- above 16Hz is a predatory sound and bellow is conspecific sound. There is no in-between, it either is a predator or isn't. Are we, then, to attribute cognition to crickets and birds?
Further, can we conclude that if a species has more categories they are more cognitively aware? If this is true, how do we compare this across species, we cannot ask a walrus how many ways they divide the world.
To say that cognition=categorization, in my opinion, is a reductionist claim. Categorization is an feature of cognition, but they are not the same.
In my tiny undergraduate opinion, categorization is an extension of working memory. Working memory is the ability to hold “on-line” many sensory modalities at once and compare them. We need working memory in order to categorize. The longer we can hold onto an experience the easier it is to start dividing the experience into parts.
It is interesting to finally have a formal explanation of categorization. I am under the impression that varying degrees of abstractness, can push categories further away from, or closer to sensorimotor grounding. To categorize a sunflower, I can look at images of sunflowers or see it in person, and receive positive reinforcement that the images that are being reflected on my retina are indeed sunflowers. In addition, I may also receive negative feedback that I have misidentified a daffodil and dandelion as sunflowers and this supervised learning would help me categorize. Sunflowers have a level of abstraction (or concreteness) that allows them to be identified easily through direct sensory experience. This does not mean that description of a sunflower through 'hearsay' would be completely insufficient in allowing me to correctly identify a sunflower without any previous direct sensorimotor interaction with one.
ReplyDeleteA concept like justice on the other hand, is in a sense further removed from direct sensory motor grounding and requires more use of ‘hearsay’ in order to avoid trial and error supervised learning. Showing me 5-10 pictures of what could be considered justice, would not have the same impression as showing me 5-10 picture of sunflowers. Before reading this article, I would have concluded that ‘justice’ could not be grounded through sensory-motor experience. ‘Hearsay’ allows us to take shortcuts through chains of abstraction. It allows categories which have previously been grounded in sensorimotor experience to be identified and chained together to create further levels of abstraction. Is it safe to say that everything is grounded in sensory motor experience but the degrees of separation from these sensory experiences is what makes some categories harder to describe and identify than others?
Categorization is cognition. Categorization requires direct sensorimotor grounding or the use of hearsay to create new categories and create degrees of abstraction. Therefore, it is safe to conclude that Cognition requires sensorimotor grounding
Nadia, all categorization is abstraction: Your brain has to abstract the invariant features that distinguish the category's members from the non-members, otherwise you don't know which is which, so you don't know what to do and not do with what (including what to call it). But, yes, I suppose the hierarchy from apple to fruit to food to thing is a hierarchy of degree of abstraction (although they are all perceptible by the senses, so in that way sense are all still "concrete.")
DeleteI wouldn't want try to ground the noun "justice" by pointing to members and non-members of the category (direct sensorimotor induction). But that could be part of the grounding of the adjective "just" vs. "unjust." (Same for truth/true/false and goodness/good/bad and beauty/beautiful/ugly.) Pointing to an example of the adjective is easier than pointing to an example of the noun. Nevertheless, the reason apes cannot learn such categories is not just their lack of interest: It's easier to explain in words what's just and unjust than to abstract it from examples. And for that you need subject/predicate propositions made out of the names of grounded categories, and having truth values -- which in turn is closely related to positive and negative feedback: The proposition "Bullying is unjust" (true) is closely related to "Bullying is just" (false). But with propositions the default assumption is that they are true. Without that, language could never have evolved.
But whereas pointing to the referent of a word can be correct or incorrect, and a proposition can be true or false, running, swimming, and even imitating cannot be true or false (though an imitation can be closer or further from what it is an imitation of: everything is similar to everything else, but to different degrees).
Lots of food for thought here...
RE: “Categorization is the problem of sorting them correctly, depending on the demands of the situation”
ReplyDeleteWouldn’t this mean that categorization is dependent on culture—A support for linguistic relativity?
According to Wittgenstein, meaning is in the “use”; and using/doing is constrained by the affordances that one’s language (situation) provides.
Although the experience of “seeing” (sensing) things may be universal, analyzing/doing/abstracting is funneled through one’s language.
Hearsay goes further by surpassing the universal sensorimotor experiences, and relying on linguistic building blocks to form new categories, for which to understand the world.
Categorization is a learned skill, where one continually refines how they react differentially to stimuli. Categorization depends on sensorimotor affordances, and “is” cognition. Understanding “how” one cognizes, therefore, depends on providing a causal mechanism for “how” one categorizes.
Therefore, “How” one categorizes is dependent on both unsupervised (sensory/innate/universal) and supervised (situation/relativity/language) learning.
Manda, to categorize is to do the right thing with the right kind of thing: What decides "right" from "wrong"? The consequences of categorizing rightly and wrongly. Nourishment, if you eat an edible mushroom; a stomach-ache if you eat an inedible one.
DeleteBut the consequences are not necessarily just physical. If you live in a dictatorship -- for example, in my native Hungary (and perhaps soon in Trump's America) whether you call abortion a "crime" depends on the prevailing culture (just as some cultures name both "blue" and "green" and some just name "bleen"). So, yes, the Whorf/Sapir Hypothesis (that your language and culture influences the way the world looks to you) has some truth to it.
Wittgenstein's "word meaning is word use" has some truth when it comes to the word "crime" or even "green," but not for "edible/toxic." And you won't be able to ground even green just from its use in the language, becauselanguage is just T2 (squiggles and squoggles). The grounding of "green" comes from its use in the world, not just its use in language. (You can learn syntax from use, because syntax is mostly just formal; semantics, i.e., meaning, isn't.)
Yes, a lot of abstracting is done verbally (through propositions: definitions, descriptions, explanations, instructions). But nonverbal animals can and do and must categorize, hence abstract, too.
Hearsay (instruction) can go further, faster than direct sensorimotor induction -- but only by recombining already grounded categories; not by its bootstraps.
One can learn to be better at categorizing, but categorizing itself is not learned: it is learning (what's the right thing to do with the right kind of thing).
Sensorimotor learning can be aither unsupervised or supervised. Verbal learning (instruction) is neither of these. It's something else.
The case for categorisation as a central part of cognition is pretty convincing and seems to be gaining popularity in cog sci (Douglas Hofstadter recently wrote a good book about it: Surfaces and Essences: Analogy as the Fuel and Fire of Thinking).
ReplyDeleteTo add a somewhat trivial point to the selective forgetting/selective ignoring story, recognising likeness seems best though of as an indifference to differences. In this sense recognising is not something positive, but rather negative: it is the failure to recognise certain key differences between two instances of similar things. As the article both explicitly and implicitly argues, categorisation is an immensely powerful tool to navigate the world, yet in some ways its the result of a kind of perceptual/cognitive laziness. (But obviously for categorisation to be effective there needs to be some specific precision weighing based on the sensorimotor system and the kinds of cues/activities that are most important for it in a given environment).
I found that part of the article super interesting. Obviously we ignore certain things when we perceive something, but how is it we're able to specifically tune out things unimportant? Or, perhaps certain things aren't more important than others, it's just how we've specifically been conditioned to ignore? Perhaps everyone has a different mechanism by which they ignore differently?
DeleteAuguste, analogy (seeing similarities) is not categorization (though it sometimes helps in discovering invariants).
DeleteCategorizing is not "recognizing likeness." It's doing the right thing with the right kind of thing: learning to detect which features are relevant and which to ignore. Word-play on detecting likenesses vs ignoring differences does not really illuminate the mechanism one way or the other. Categorizing is not a tool: being able to categorize is a tool. But it's not a lazy tool: your brain (if not you) has to learn to detect the invariants. Cogsci is looking for the mechanism. ("Deep learning" neural nets are the latest candidate.)
Laura, which features of a thing are invariant within the category, and distinguish it from another category, depend completely on the thing, and on what you need to do with it (and not-do with another kind of thing).
We don't yet know the mechanism of category learning, but it's certainly not "conditioning" (whatever that means: Pavolvian "classical" conditioning?) It's some form of supervised learning, with corrective feedback.
The ugly-duckling story made me think about what exactly constitutes a category? I like the notions of common features, or for visual categorization geons. However, how do these hold up in the cases of family resemblance? Family Resemblance is when each of the members of a category share at least one feature with another member of the category, but there is no feature which they all share. In the paper ‘underdeterminism’ was discussed and it seems as if, some of the ways we categorize when we cannot explicitly say, may have to do with this notion of family resemblance.
DeleteValentina: Family resemblances (Feature A & B, or B & C, or C & D, or D & A) are still features, and can still have feature detectors for the invariant feature, which is the entire Boolean combination rule. (Don't trust everything you hear from or about Wittgenstein, a giant, or Rosch, a pygmy!)
DeleteRe: Abstraction and Hearsay
ReplyDeleteThis section explains how we go from abstracting to naming things, with the intermediate step of recognition. Our sensorimotor systems have the capacity to not only passively react to stimuli but to also be active. To abstract is to single out some subset of the sensory input and to ignore the rest. For instance, if I were looking for pineapples in a fruit salad, I would isolate what I believe to be pineapples and ignore the rest of the variation. We have the capability of detecting such invariants selectively but we still need to find the mechanism behind it. Selective abstraction requires knowledge of alternatives among which the isolated object needs to be sorted. We compare the invariance relative to the variances around. The next step is recognition which is more than just seeing. This is an active step where we see something as a kind of thing that we have already seen before. Recognition is built on experience, drawing on feedback and giving salience to certain features of a stimulus. Once we have recognized the stimulus in question, we can proceed to naming it which takes us further into the process of categorization. This paper argues that cognition is categorization. In going from abstraction, recognition, and eventually to naming, do these processes occur discretely or continuously? What process is responsible, if any, for our ability to abstract invariants from variants? What is the causal mechanism for categorization?
Neil, we don't know the mchanism of category learning yet (though there are some candidates). It's learning to abstract and detect features and then works like a kind of filter, ignoring the irrelevant features and using the invariant ones to do the right thing with the right kind of thing. There is a big jump from sensorimotor induction to verbal instruction.
DeleteRegarding Funes the Memorious:
ReplyDeleteIn the article, it claimed that Funes would be unable to understand why anything belongs to any category, whether it be the same object in different points in time or why two four legged, fur-covered mammals that are different sizes and colors are both called “dogs”. My understanding is that this is because he is unable to “ignore” the differences between the stimuli, and therefore cannot draw a category around them both to designate that they are the similar. This brought me to the question: If Funes were to look at a spectrum of the frequencies of light (color spectrum), would his just-noticeable difference be the exact same as if he were to look at a spectrum from light gray to dark gray? My initial assumption would be that they are the same, but with that in mind, wouldn’t every single pixel be its own category? Funes would certainly be able to tell us that two things are different, but that is limited by his perception in the same way that we are limited by our perception.
Karl, Funes could make same/different judgments (if he could understand the instructions -- which he could not, because he could not have language, because he has no categories) as long as both stimuli were still present. How small a difference he could perceive is a question about the size of his JNDs (just-noticeable-differences), which, since he is a fiction, is not really relevant.
DeleteHarnad strongly rejects the innate aspect of categorization and says it is all evolution or learning. For humans, we do see inborn features (perhaps as a result of evolution as Harnad would suggest) such as fallible memories to allow selective forgetting, and primary color and speech categories and these give certain affordances that allow us to percept shadows in a certain way. How would these inborn features be applied to a robot? It seems to me that having a machine to only use supervised learning to learn to apply these inborn features would be incredibly slow in terms of learning to apply these in all the instances humans naturally do, and to do it as good as we can. Humans do an enormous amount of unsupervised learning and are constantly passively storing and updating this map of the world and context in our memory. Thus, it seems like selective forgetting in particular and unsupervised learning would be an incredible challenge for machines to do as well as they are able to do supervised learning. Would it be possible for a machine to simulate evolution over generations (obviously this would be much faster for a machine to do) and to allow them to eventually establish weights and inputs that mimic these inborn features? Is this mainly a computational challenge or one that should be considered when setting up sensorimotor capacities in a T3 and perhaps biasing them to reflect these natural features of human cognition?
ReplyDeleteJulie. not all categories are learned: most are.
DeleteWhat's inborn in a human can be in-built in a robot. Yes, evolution can be simulated too, in the robot, but why? For innate categories we need only the end product, not its real-time history. And to simulate evolution, you also need to simulte the evolutionary environement (i.e., the world, or at least the organism's niche).
Both robots and humans can have inborn categories, supervised learning and unsupervised learning. Dynamic processes, too, can be simulated computationally, but as long as they are purely internal, and work, it doesn't matter.
In addition, evolution in a robotic/computational setting can be much much faster and efficient than human evolution. For my thesis right now I'm doing computational modeling of evolution, and computers make it possible to carry out experiments that represents thousands and thousands of years over the course of a few minutes (or less!). This means that if we can create robots that are good at learning over time, the "time" aspect that could take years (or generations) for humans could be almost instantaneous.
DeleteSo while for categorization, we might not even need to simulate any sort of evolution, if there are other tasks that involve changing behaviors, knowledge, or understanding over time, this could potentially be carried out much more efficiently in a robot than a human.
Dominique: You would be biassed in favor of robots, wouldn't you...
DeleteBut remember that cogsci is about reverse-engineering human capacity rather than producing useful slaves for humans...
Re: “To be able to abstract the shared features, we need supervised categorization training … the inputs are recoded … the features are re-weighted.”
ReplyDeleteFrom what I understand, our sensory systems weigh certain features more heavily than others, which is equal to saying it abstracts certain features as privileged. Relating this to how categories “consist in ways we behave differently toward different kinds of things”, it looks like feature selection and weighting depend on things’ affordance. But here I’m suspicious; is affordance a weasel word? Isn’t it just about practicality – i.e. what I use things for. Objects can afford certain sensorimotor interactions, but what underlies this, in my view, is what we do with these things for (i.e. for what do we do with these things?). That is to say that we don’t behave differently towards different kinds unless the kinds differ in their practical value for us. Does it follow that categorization is fundamentally based on practical value for us and that practicality is what explains feature selection and weighting? I argue yes. Instead of what objects afford, it’s what I do with objects that matters. If this first point is defensible that affordance is just a weasel word for a thing’s practical use for us, then categorization faces the problem that Eleanor Rosch described: we cannot state the invariant basis on which we categorize. A table can “afford”/be used to sit on just as a chair can. We can (almost) use any objects for any use. Thus, it seems that family resemblance is separate from the classical theory of categorization. There are uses which are more similar and thereby, form a ‘family’. Family resemblance serves to exhibit the lack of boundaries and the distance from exactness that characterize different uses of the same object. All of this is not to say we can’t categorize. But I argue that the feature selection and weighting involved are not about invariant features; it’s about travelling with an object’s uses through “a complicated network of similarities overlapping and criss-crossing” (Wittgenstein, Philosophical Investigations).
Austin, doing the "right thing" depends on "practical" value (survival, success), but it's not all based on sensorimotor-affordances. Invariants can be just sensory too.
DeleteWe have an idea how feature abstraction with trial-and-error and feedback (supervised learning) might occur. There are models for it, that work. No one has any idea how " travelling with an object’s uses through a complicated network of similarities overlapping and criss-crossing” might occur (or even what it means).
No, neither (1) categorization (doing the right thing with the right kind of thing) nor (2) affordance (sensorimotor invariants that depend not only on the "shape" of the object but the shape of our bodies is a weasel word.
Yes, mushrooms can be said to "afford" being categorized as edible or inedible as a result of trial/error learning and feedback that abstracts their invariants -- but why call those invariants "affordances" if they are just sensory?
Following up on our class discussion on what it means to do the right or wrong thing with different kinds of things, I think I have grasped the all or none logic. I went back to the reading and read the rebuttal of family resemblance again. It seems that I misunderstood or simply glossed over how family resemblance is simply disjunctive invariants. Where I still have some issue is that these disjunctive invariants vary depending on objects affordances. What we can extract from our motor interactions with our sensory input is astronomical. Objects afford as many things as we can do with them; doing the right or wrong thing varies to a large degree. Supporting this, when we consider the meaning of right/wrong in terms of consequences, consequences vary depending on a multitude of factors, not only evolutionary/survival reasons. Thus, despite weighting and selection, wouldn’t the disjunctive Boolean phrases be astronomical in its complexity as well? Furthermore, it seems that the disjunctive Boolean phrases would vary quite a lot since right/wrong varies as well. So, while the disjunctive Boolean phrases per se offer an all or none boundary (which I now see), there doesn’t seem to be much meaning to constantly moving boundaries.
DeleteAn example of where doing the right/wrong thing with different kinds of things doesn’t depend on survival/evolution is how we categorize Jaffa cakes as discussed in a recent BBC news article:
Deletehttp://www.bbc.com/news/magazine-38985820
"This kind of extreme nativism about categories is usually not far away from something even more extreme than nativism, which is the view that our categories were not even "learned" through evolutionary adaptation: The capacity to categorize comes somehow prestructured in our brains in the same way that the structure of the carbon atom came prestructured from the Big Bang, without needing anything like "learning" to shape it.
ReplyDelete(Fodor's might well be dubbed a "Big Bang" theory of the origin of our categorization capacity.)"
Surely it's a mistake to say that all categories are innate, but it does seem like at least some of them are. Even beyond UG, there are physical things in the world that we seem to know about from birth, and especially to fear, despite the impossibility of having ever encountered them in a context of imminent danger, or at all.
Many of us are inherently afraid of snakes and spiders, even in Canada where venomous creepy crawlies in natural environments are few and far between. Even cats seem to have an innate fear of snake-shaped things (https://www.youtube.com/watch?v=cNycdfFEgBc), despite having never seen a snake nor even heard about one.
Many children are afraid of monsters lurking in their closets or under their bed. Is this a case of wild imagination, or are "monsters" really an unconscious analogue for deadly predators like bears, wolves, big cats, and the like?
These sorts of evolutionarily selected-for categories are geared towards survival, and are clearly advantageous in an environment where things that can kill you may always be right around the corner or under your feet. I can see a case for these sorts of categories being innate, but it seems obvious that kinds we would never encounter in an ancestral environment - kinds like 'computer', 'boomerang', or 'Jerry Fodor' - cannot be inborn.
Michael: Yes, there are definitely other innate categories besides UG (likes snakes, spiders, heights, the dark -- and especially the sexual ones, which we will be discussing next week...) But the point was that most categories are learned. And that even innate categories have to have invariants, and evolved invariance-detectors.
DeleteIt seems that categories are either discrete or continuous. The continuous case, however, often has cultural norms embedded within it. Another key aspect to continuous categories, however, is relativity. As mentioned in the reading, antonyms like “hot” and “cold” are similar to colour categories. The words “hot” and “cold” may be somewhat discrete because most humans can decide on the threshold of where something is hot verses cold but “warm” and “cool” can only be decided relative to one another. I would argue that “true” and “false” are also discrete while something like “beauty” and “ugliness” (in the ugly duckling case) are extremely relative and subjective (and culturally influenced). Discovering these boundaries could potentially be done by observing speech and finding common collocations with certain antonyms or objects. The real question, however, is what are the properties that make these categories discrete vs continuous and how do we as cognizing humans pick out whether an object or adjective falls on a spectrum or is its own entity?
ReplyDeleteOn another note, the following statement from the article, “What the stories of Funes and S show is that living in the world requires the capacity to detect recurrences, and that that in turn requires the capacity to forget or at least ignore what makes every instant infinitely unique, and hence incapable of exactly recurring,” nicely sums up how adaptive our memories are. This ties back to previous discussions we’ve had about a T3 robot not resembling a human because its memories are virtually limitless. Categorization, however, may be the key to programming a computer with the ability to cognize as humans do, in a way that is inherently adaptive.
Annabel, I think even with culture aside (although probably with greater variation between cultures) we have this discrete vs. continuous dilemma you mention. To me this problem identifies a merky boundary between arts and sciences. I think the arts concern more the subjective, continuous, and the sciences concern more so the objective and discrete. But neither seem mutually exclusive. In consideration of your question, in my opinion the properties that identify the discrete vs. the continuous are arbitrary when it comes to distinguishing which words/concepts pertain to which category. Some categories like red/blue are more objectively distinguishable (discrete), yet can be more ambiguous for the colour blind. Conversely, I think some categories are more relative (continuous) like higher rates of schizophrenia in cities, but can be ambiguous when we consider the elements which make a city a city (e.g. lack of green space, mass production). But both kinds of thinking exist in the same world, and distinguishing one from the other seems impossible when both what we are calling “continuous” and “discrete” have their foot in the other category. We are on shaky grounds of categories within categories!
Delete**rates of schizophrenia may not be a very accurate example. Perhaps experience of mental illness would be more descriptive of a relative category
Deleteo Never thought I would get the hang of these abstract conversations. Very well explained concepts this week! I just had a couple of questions. Categorical perception (CP) was described as an innate process; a categorization process because the learning (i.e. trial and error; feedback process) was Darwinian. It is very interesting to frame natural selection as a feedback process in which a lack of CP, or perceptual distinctions led to dying out. I understand that in computer modelling, learning processes that embody feedback and trial and error are executed and sophisticated. I wonder, however, if the threat of extinction (presumably a powerful force) such in the process of natural selection is a key ingredient in categorization learning that is missing from computer modelling. I suppose individuals are not necessarily aware of the natural selection forces that act on their fitness, however it may on some level motivate their system to adapt and survive. Secondly, when discussing all the possible permutation of invariability, it is stated that “if we return to sensory inputs, and the problem facing the theorist trying to explain how sensorimotor systems can do what they do, then sensory inputs are the shadows of a potentially infinite number of different kinds of things. Categorization is the problem of sorting them correctly, depending on the demands of the situation”. It then says that as long as unsupervised learning is underdetermined and not indeterminate, and does not lack poverty of stimulus, it can help. Is this how we think humans categorize? Or just what is needed for modelling in which case even if it deviates from the way humans do it doesn’t matter?
ReplyDeleteIn regards to your inquiry about computer modeling accounting for a 'threat of extinction' or an akin concept, my 2 cents are that computational simulations for game-playing or an agent reaching a particular situational goal do in fact assign appropriate weights to and account for outcomes that are negative or less optimal. For instance, various Monte Carlo algorithms select what appears to be a good action and then averages out the evaluation of repeatedly sampling a continuation from that action. In doing so, consequent 'errors' or not so optimal consequences are taken into account.
DeleteRegarding the Funes example: there is something extremely economical with having categories.
ReplyDeletethe same way we could have all of our plates, utensils, mugs pots and pans
distributed randomly all around our kitchen, it is much more efficient
to have drawers, cupboards and shelves.
I see the brain the same way. Having categories allow us to put
things into “mental boxes, and therefore access concepts much more quickly.
When we need to access our “spoons” we know which box to open and can
access the info we want a lot quicker. Considering our brain limited capacity, this ability is definitely crucial.
I dont want to get into grammar too much, but it seems clear to me that there is a parallel between learning a language and learning a category: in learning a language, there is also a type of positive reinforcement (listening to people speak correctly) and negative reinforcement (being corrected when we make a mistake), that seem similar to me with supervised and unsupervised learning. However in language acquisition, it seems like negative reinforcement is not a necessary condition of learning a language, for example, children who learn 2 languages from birth tend to start speaking one language earlier than the other, and start speaking the second language later (around 5 or so), with a very high level of complexity (considering the fact that they have barely said anything in that second language yet and have therefore not received almost any negative reinforcement, only learnt from listening). Would you agree with the parallel I am drawing here? if so, would it be a basis to hypothesize that supervised categorization is useful but not necessary?
ReplyDeleteI agree with the comparison you are making but wonder if there actually is much negative reinforcement when learning a language. It's very rare (and seen as rude in most cultures) to correct someone who makes a grammatical error in speaking, so most of language learning is done by what you call positive reinforcement.
Delete"... the adaptive value of language: Language allows us to acquire new categories indirectly, through "hearsay," without having to go through the time-consuming and risky process of direct trial-and-error learning. Someone who already knows can just tell me the features of an X that will allow me to recognize it as an X."
ReplyDeleteI really like the above example & think there is strong support for the value of language for categorization. However, I wonder how this affects animals that do not communicate (or communicate in a less complex way than humans). I would assume that animals form categories just like us, as it's an adaptive trait for survival. Does this mean, then, that animals would form categories at a less efficient rate? Or is the formation of complex categories less important for certain animals compared to humans? This would be my hypothesis (maybe this is why language may not have developed for them?), but I'd love to hear others' ideas.
I had a similar thought with regards to animals on this quote as well.
Delete“Yet what explicit knowledge we do have, we can convey to one another much more efficiently by hearsay than if we had to learn it all the hard way, through trial-and-error experience. This is what gave language the powerful adaptive advantage that it had for our species.”
I agree with this statement about language giving humans this adaptive advantage. I am wondering if other mammals who communicate (whales, primates) in their own ‘languages’ or means of communication also share this adaptive advantage? While our language is more articulate and therefore can be argued to be superior, can sounds and gestures mutually understood between other mammals not convey the same power? When one thinks of sign language, gestures that represent words/language and that are based on said language, it is evident that signing too has the equal power that spoken language has. Could one argue that animal gestures not based on a verbal/spoken language hold the same or any weight?
Comment on Explicit Learning:
ReplyDelete“Biederman did a computer-analysis of newborn chick-abdomens and identified the winning invariants described in terms of his “geon” features (Biederman & Shiffrar 1987). He was then able to teach the features and rules through explicit instruction to a sample of novices so that within a short time they were able to sex chicks at the brown-belt level, if not the black belt level. This progress should have taken them months of supervised trial-and-error training, according to the grandmasters.”
I am wondering what the implications of thinking of categorizing as abstracting some features and ignoring others are? If we categorize by reducing to geons and focusing on certain elements while disregarding others, where does the idea of a template come in? If this process is truly innate and/or based on corrective feedback, how do templates of categories come into being and achieve any form of status?
RE: Feature Selection and Weighting
ReplyDelete"We can see, while they are present, far more features than we can remember afterward."
To say "we can see" (features) is already complicated by the neurological correlates of feature selection, as neither the eye nor the brain truly sees anything. In fact, although the photons hitting our retinas reflect an approximation of what lies before us, as you said, our sensorimotor systems don't even detect all features.
Feedforward circuits of perception, in isolation, still already have dropped certain details of the physical reality that we will never even encode for, much less forget about later.
These may as well be adaptive shortcuts however, as the number of feathers on the chicken is biologically unimportant to the human observer.
Even the hypothetical Funes, with perfect recall, lacks the capacity to perceive much of the world.
Much like last week this article reminds me of Bertrand Russell's work in "The Problem of Philosophy". Specifically his thoughts on categorization and abstraction levels. This article gave me more of a depth of understanding in the mechanisms and functionality of categorization in cognition. I would like to focus on the concept of Deep Learning.
ReplyDeleteIn the link you provide, it mentions that "Deep learning discovers intricate structure in large data sets by using the backpropagation algorithm to indicate how a machine should change its internal parameters that are used to compute the representation in each layer from the representation in the previous layer." An example of this could be our visual cortex, V1-V6, correct? When point representations of an image move up in abstraction as the information is passed onto higher V levels (such as line representations to 3D shapes based on spatial positioning and color, etc). Are such deep convolutional nets able to self regulate, such as out own sight? For example, when we observe scene in a movie, we are able to discern it as a movie and not as a scene seen through a window (happening in real life)?
One line in the 'Supervised Learning Tasks' section was particularly interesting to me and was not elaborated on.
ReplyDelete"Note that this presupposes such thing as an error, or miscategorization".
This is not brought up again in the rest of the paper, but I believe it poses some very interesting questions. Are there technically correct or incorrect categories? The innate categories could I suppose be classified as correct: it is incorrect for someone who is not colour blind to say that a red object looks blue. For other categories, however, this distinction is less clear. A later section discusses how certain features are augmented to form categorical distinctions, as in the case of the swan and the ducklings. Would there conceivably be a correct and incorrect way to augment details, are some intrinsically more pertinent than others, or is it all based on the distinctions we were taught to attend to?
You bring up some really interesting points! I agree that for innate categories there could be objective errors and correct categorization. However, I'm wondering why even in people who can perceive color there is such a difference in categorization? For example, someone may think a color is teal while someone else thinks that it is blue. Is this an error in learned categorization of where labels fall on spectrums of color? Is this a difference in ability to perceive small variations or details of an object?
DeleteI think learned types of categorization could also be considered correct or incorrect, but on a subjective level. Learned categorizations are more influenced by environment and culture, and so then it would be that environment and culture that would determine if things are being correctly categorized. It is the other people around you enforcing these categories. For example, in a city like Montreal, there may be one conception of injustice (as a more abstract category, mentioned elsewhere in the comment section), whereas that same occurrence would not fit into the category of injustice that is learned and upheld in a small rural town.
One part of this article that drew my particular interest was the idea (in section 21. Recording Feature Selection) that to increase our capacity to make categorizations, we can add more dimensions of variation... "but even higher dimensionality has its limits."
ReplyDeleteFirst, I'm wondering if this idea has been further looked into, either through research or real-world examples. This idea could be used to create tools to better categorize or remember aspects of the world (perhaps allowing everyone to surpass the 7+/- rule instead of only those with synaesthesia). Moreover, this seems like an important key to machine learning, and seems to point to one reason as to why machines need T3 (sensorimotor capability) to pass T2 (linguistic capability). In order to make competent categorizations, machines would need as many sensorimotor variations and distinctions as human beings have.
RE: In machine learning there is a “credit assignment problem” (CAP) which poses the question of: “how do we find the rule amongst many possibilities?” (Harnad 2005 referencing Sutton).
ReplyDeleteAs I understand from the paper, CAP entails pinpointing some conformant plan which would allow one to distinguish between multiple entities based on a discrete feature(s). Harnad continues to explain that the CAP may be addressed by a hybrid unsupervised and supervised learning schema - the former allowing for multiple exposures and self-discrimination and the latter providing feedback following trial and error. The CAP relates to chunking in that 'chunking' together several distinct features allows for the recognition of a more cohesive entity, e.g. a face can be categorized as being comprised of two eyes, a nose in the center, and a mouth, all arranged in a specific way). The CAP might pose the question of distinguishing between a cat and a dog - both are furry, have faces and tails. There would have to be some pattern specific to cats and another specific to dogs such that cats can be categorically distinct from dogs and vice versa. The paper proposes that a combination of supervised learning - characterized by some feedback following trial and error - and unsupervised learning which allows someone to undergo repeated exposure and ultimately self-discrimination may be a solution the CAP.
Chunking in computational modeling is particularly fascinating with respect to the CHREST model for long term memory where lateral links are created between nodes representing a state or information (simulating chunking) and for consequently developing more "complex schemata" representative of semantic memory (Gobet et al, 2001).
Fodor has argued that there is the “vanishing intersections” problem, so that is why categories are innate rather than learned or evolved. However, I found the examples given in this argument relatively weak. Saying that there is empty intersection among “goodness”, “truth” or “beauty” is not very convincing. I think there are definitely learned categories by association or, “intersections”, in this case. For example, words in a similar category like “potato” and “sweet potato”, or words like “treat”, “treatable”, “untreatable”, have intersections not only shown by their “physical” similarity but also associated syntactic meaning.
ReplyDeleteI agree that the “Vanishing Intersections” argument is weak in relation to concepts like “goodness”, “truth” or “beauty”. Fodor believes that the intersection, the overlapping features shared between category members, is empty since these concepts are more abstract and (upon first glance) seem to lack shared sensory, or even functional, features. While I agree with you for the most part, I’m confused by intersections shown by “associated syntactic meaning”. I’m not sure how Fodor’s argument concerning conceptual categories can apply to syntax, since Fodor is more concerned with meaning, and based on my understanding of previous lectures, we cannot derive word meaning from arbitrary symbols or syntax.
DeleteFrom my understanding, Fodor assumes the innateness of category learning from generalizing Chomsky’s reasoning for Universal Grammar: that we cannot learn categories via trial-and-error and corrective feedback after being exposed to however many examples since there is no commonality between these examples – i.e. the intersection is empty. However, even when looking at abstract concepts, like the ones aforementioned, there is always some overlap between the concepts, otherwise there would be no meaning to the concepts at all. It’s for this reason that vanishing-intersections argument is not correct in saying that category learning is innate, since although we cannot conclude from its immediate sensory shadows that the examples share a common feature, it is not to say that a commonality does not exist.
Yea so after class I realized that most of categorizations are learned and are not innate. However, theories about innate categories can be right like Chomsky’s similar conjecture about universal grammar. Also like our predilection for sugar and fear of spiders that we covered in next week’s class demonstrate the learning of categories through evolution.
DeleteInvariant Sensorimotor Features ("Affordances")
ReplyDeleteHow about blind spot or illusions? We don’t really have the correct sensorimotor or the sensory apparatus, yet we have visual processing which compensates for the lack of the apparatus, which inevitably leads to false processing such as illusions (Ebbinghaus illusion, Ponzo illusion, etc).
Autonomous, adaptive sensorimotor systems categorize when they respond differentially to different kinds of input...categorization is intimately tied to learning.
Here’s an example. Imagine two identical human beings. They have the exact same physical composition and structure (identical to the molecular level). If they were exposed to the same experience, would they respond the same to a question? For example, if they were trained as doctors, would they produce the same diagnosis for an illness? How about ethical decisions? If this experiment demonstrates that same experience and therefore learning, leads the same response, the only reason why human beings are unique is due to the differential experience we all are exposed to.
Learned Categorical Perception and the Whorf Hypothesis. & Whorf (1956) suggested that how objects look to us depends on how we sort and name them.
While thinking about categorization, isn’t categorization grounded in creativity? Creativity is not creating something new, but simply the ability to mix/combine old information. Also, creativity is the ability to place old information in a different context (place already categorized category into another category). For example, BROOM; a tool used to sweep the floor. That’s what we would think. But give that broom to children. They will play sword fight, gunfight, it will even become the vehicle for them to fly on (i.e., Harry Potter). Why are children able to find a new category that we conventionally are not able to identify? Fundamentally, children strive to have fun. Their motivation in everything lies in having fun. Not to be confused with hedonism, “fun”-ism relies on the fun aspect of all things; it is not mere indulgences.
On your first point, I believe most of the major theories regarding blindsight claim there is some intact sensory apparatus remaining, whether it be residual V1 neurons or alternative pathways bypassing V1 entirely. Thus, the features are still extracted, but they are not available for conscious report.
DeleteThe experiment you mention as your second point is similar to a question asked by Professor Harnad a few weeks ago, the question of why all students in the same class, given the same information, do not understand the material with equal clarity. If I am correct in understanding these thought experiments, they hint at if there is anything more to the mind than the physical hardware (the brain) and the inputs/outputs. If your two identical beings would not produce the same output (ex. diagnosis) given the same inputs, with the same internal processing, then it would indicate there is some non-physical (or currently undetectable) phenomenon governing cognition.
From: The Adaptive Advantage of Language: Hearsay.
ReplyDeleteHarnad mentions that sensorimotor learning often occurs implicitly, which makes it difficult for one to explain correctly through language. However, he also states that language has allowed us to convey the explicit knowledge we have much more efficiently than if it were solely done through trial-and-error. With that being said, in order to have complete categorical understanding, can we say that language as well as sensorimotor experience are but necessary but neither is entirely sufficient on its own to reach this state? To add to that, in order to replicate CP in a machine it would have to be at the level of T3 to have sensorimotor modalities as well as linguistic properties and symbol grounding specifically.
RE: They have suggested that one of the reasons most categories can be neither learned nor evolved is the “vanishing intersections” problem: If you go back to the dictionary again, pick some content words, and then look for the “invariance” shared by all the sensory shadows of just about any of the things designated by those words, you will find there is none: their “intersection” is empty.
ReplyDeleteJudging from this argument, it seems fair to say that grammatical rules are inborn - they are just rules that you can and cannot say. But what about meanings? Are meanings inborn? Meanings are not simply just rules that society has shaped. As we grow up, categorisation provides mental representations of meanings that are mapped to words. Since linguistic labels help categorisation - these two learning processes interplay with each other to develop the feedback loop of trial and error. Thus if grammatical rules are inborn, and so are categories, what about meaning?
I find the concluding statement – "all of our categories consist in ways we behave differently toward different kinds of things... And isn’t that all that cognition is for -- and about?" – a bit challenging in respect to some of the other points made in the course. Isn’t categorization something which can be done by a symbol manipulation computer? We can program Turing machines to categorize with accuracy approaching our own. Assuming a more advanced level of technology, a computer should be able to perform differentiation behavior (visual discrimination, for example) to the same degree that we can.
ReplyDeleteIf ‘all cognition is’ is learning categories and how to behave towards it, and Turing machines can do it, doesn’t this challenge the notion that digital computers cannot do cognition?
Clearly cognition is intimately related to categorization, but how can we maintain that computers =/= cognition if they can do categorization in the ways that we do?
These readings, and the idea that cognition is categorization, have made me interested in the links between memory and categorization. As demonstrated by the “Funes the Memorious” story, memory is involved in categorization in the sense that a person must be able to ignore/forget invariant features in order to focus on the variant ones. Additionally, memory is affected by the ability to chunk information (e.g. grouping binary 0s and 1s in threes to represent numbers), thus increasing the number of single digits that can be remembered. Can these sets of three be thought of as objects and the digits as features of those objects, in the sense that they make it much easier to remember the features of objects around us when we know what “kind” of objects they are, as opposed to attempting to remember each feature of each object individually? If so, this seems like somewhat of a circular issue— too much memory ability impairs the ability to categorize, and yet one of the benefits of categorization is that it allows us to remember more information.
ReplyDeleteAdditionally, based on these two examples, memory and categorization appear interrelated. However, does categorization help in all memory tasks, or is memory one aspect of cognition that is not merely categorization? Let’s look at the example of remembering the name of a third grade teacher. Is it possible that when we remember something, we do so by remembering categories, for instance teachers, then specifically that teacher, and that her name is remembered as a “feature” of that chunk of information, in the same way that digits can be remembered by first recalling the number they represent?
Emma, I too found the relationship between memory and categorization the most interesting in this reading! My fascination with memory is that it seems to both limit categorization and enhance it.
DeleteAccording to Miller, who says that there are only 7 regions along a dimension for which we can categorize, if we had more than 7 regions we could theoretically categorize more.
On the other hand, the Funes story shows us that an infinite memory is a hindrance for categorization because we need selective forgetting to distribute weights to affordances. With a perfect memory, each infinitesimal affordance would receive an equal weight, preventing categorization.
Considering both sides, I wonder if there is an optimal amount of memory for categorization that characterizes conscious machines. With zero memory we become simple dynamic systems like wind blowing on sand in the desert. With too much memory, do we become like a computer that cannot categorize with the information it has without an external mind providing algorithms for it? I'd appreciate anyone's input that could clarify the possibility that memory can both aid and hinder categorization.
Also, I think we do remember something like our 3rd grade teacher by remembering categories. I found myself first thinking of the classroom associate with "3rd grade" and then I look up to see my teacher standing in front of the class. Unfortunately, I use an introspection argument because that's the best I can do for now...
Hey guys!
DeleteI definitely agree that categorization is beneficial to memory because it expands our memory capacity by allowing us to chunk similar pieces of information together. With regards to the "Fune the Memorious" example of infinite memory capacity, the main point that I took from this was that having an infinite number of categories defeats the purpose of having categories at all. If you have an unlimited memory, then you also have the ability to pay attention to an unlimited amount of variant features. Thus, you technically have the ability to create an infinite number of categories for these features. But this is completely counterproductive. The fact that each instant is incomparable for Fune, means that everything he experiences in his environment is unique, and technically has a variant feature of its own. Essentially, categories are completely useless to him - if their primary function in relation to memory is to help us increase our memory capacity, then someone with an unlimited memory can't make use of them. However, I do think that this example is really interesting in terms of thinking about how much memory or storage a machine should have to help it perform optimally.
ReplyDeleteRe: Is there any way to increase our capacity to make categorizations? One way is to add more dimensions of variation; presumably this is one of the ways in which S’s synesthesia helped him.
I wonder if novel, abstract and emotional experiences share similar categorization problems (if it is one at all) as does the synesthetic. If an experience is reported as metaphorically “felt”, it may accompany an unidentifiable cause or feeling rather than referred to a concrete symbol. For example, in patients with panic disorder, triggers of a panic attack are (mostly) unidentifiable. In non-pathological cases such as meditation, people report “out-of-body experiences” which are physically impossible. In these cases, the words used to describe the experience are insufficient for explaining the cause or “essence” of that experience. I wonder if this is a problem of lacking discrete categories to identify the feeling with, or if some kinds of feelings are less suitable for categorizing. The article suggests that synesthetics have an increased capacity to categorize, but could the opposite be true, that they lack the means to associate experience with respective categories?
Re: Those of our ancestors who could make rapid, accurate distinctions based on color out-survived and out-reproduced those who could not. That natural selection served as the “error-correcting” feedback on the genetic trial-and-error variation.
I’d agree that the ability to distinguish between figure and ground, to make distinctions develop out of survival necessity. This ability seems fundamental to the experience of being human (not to say all other animals are excluded!). In Funes’ case who lost the ability to forget, and thus the ability to categorize, his capability became a debilitation. But as the article emphasizes, if it could not have been possible for Funes to speak as as word use necessitates categorical selection. This confirms that part of what may make our species unique is the ability to categorize so as to make language acquisition possible.
This comment has been removed by the author.
ReplyDeleteThe essay puts forth a strong argument against the notion that categorization is innate but I was wondering what could be the explanation of the origins of selective weighting. This question is with regards to the example about ugly ducklings and the relative weighting of certain features over others which allows one to say that one entity out of the group is distinctly different from the others (although in another context it could be more designated differently). How do we learn which features to give greater importance to when categorizing objects? For example, my friends and I constantly debate what is a fruit and what is a vegetable. Intuitively we seem to know, giving priority to taste or appearance or similarity to other items in the respective categories. But then someone comes out and says than an eggplant is a fruit because it has seeds on the inside. To somebody who doesn't know the formal taxonomy, how do we give greater importance to certain commonalities between eggplant and broccoli and not, say, apples?
ReplyDeleteI think it's important to distinguish between categories that are generated as a result of experts coming to a consensus and categories that denote things from other things. So what are fruits and what are vegetables are categories of categories that were built from history, society, and science, but what makes apples apples is a more subtle and interesting point I think. One of the differences is that selective weighing is very implicit, but taxonomies are very explicit.
DeleteI think that the implicit nature of feature salience is a good place to bring forward how culture, socializing, and human experience come into play. For example, from ethnograhies studying the communicative behaviour and journals of Inuits living in the arctic, it appears that they do not make categorizes of things visually (as we dominantly do ie. colour, size, location), but do discriminate qualities of sounds by the wind. The argument is that living in a visually featureless landscape (ie. blinding white arctic landscape) led to dominance of audio information over visual to determine what is important to categories.
I find this article very pertinent to computer modelling and AI, if cognizing is categorization then researchers working on feature discrimination, object detection etc maybe closer to human computation then at first glance. The most successful models (to my admittingly limited understanding) are deep learning nets that use back propagation type learning to change internal configurations based on comparing outputs to a "correct" output vector. This seems like a simulation of trial-and-error supervised learning, which as shown in the article we do use categorize. The unsupervised learning that occurs in deep learning is very interesting, after the supervised learning occurs right and wrong can be determined, but grounding these discriminations is more complex I think. Without work from robotics, the deep learning models cannot ground the categorizations in sensory-motor interactions. If it is such a necessity to CP, thus cognition, then it's a pretty strong argument for robotics and computer modelling to be working together (ie. T3 is right, not T2).
ReplyDeleteRE: Cognition is Categorization
ReplyDeleteIf cognition is categorization, and categorization is differential (same kind of input yields the same output), in what way is mental rotation categorization? Experiments correlating the time that it takes to mentally rotate an image with the amount by which the image has to be rotated suggest that mental rotation is a continuous or analog activity. Is mental rotation (or mental imagery in general) not a facet of cognition? Maybe space and time are themselves differential (as in Planck time or Planck length).
RE: Degrees of Abstraction
Though beauty and apple are both abstract words that could refer to an infinite number of possible referents, I would argue that beauty is in fact more abstract than apple because the word beauty has a wider variance among its possible referents than the word apple does.
Gus, mental rotation is not categorization. It's a relative judment, comparing one thing to the other to say if they are the same or different. (If you name every pair of things you ever see that are the same, "same," and that are different, "different," that would be a categorization, but in a trivial sense.) If you try to copy something, and you do a poor job, you have not miscategorized it; you've just done a poor job of (analog) copying: a poor approximation.)
DeleteNot all cognition is categorization. Continuous motor skills (walking, swimming, singing, tennis, golf) are not categorization. (But Planck time and Planck length have absolutely nothing to do with it!)
Yes. all categories are abstract, so abstractness is just a matter of degree. (Only Funes's unique instants are not abstract.) But some categories are more abstract than others, because abstraction can be hierarchical (apples - fruits - foods... things), and also because some categories are necessarily combinatory, because they are based on verbal descriptions (e.g., like "mammal" or "bachelor" or "unreliable": you can point to members and non-members but can you learn the invariants by induction? or only via instruction?).
Does this mean all adjectives are categories (that are quite high in terms of their degree of abstractness)?
DeleteAlso I think the question of whether or not you can learn a category by induction or only by instruction is a hard one to answer definitively. There's many categories that typically need to be taught explicitly to children and yet it's possible that a particularly observant child might notice a "pattern" between different objects and thus conclude using inductive logic that this must be a category (although they would not necessarily have the right "name" for this category without some form of teaching).
Language affords us common labels to identify these categories in order to communicate with each other.
ReplyDeleteRE: To Cognize is to Categorize: Cognition is Categorization
Categorization obviously plays an extremely important role in how we perceive the world around us. There is a lot of evidence that suggests that our memories are formations of individual pieces of information encoded into categorical patterns across the brain. A major issue when it comes to forgetting information or confabulating events (a surprisingly common occurrence in our minds) is this incomplete pattern separation. To elaborate, one piece of categorical information could resemble a piece from a separate memory. When the original memory is recalled all of the correct components of the memory are there, except this categorical information that resembled the original one too closely. I think it’s interesting that how categorization can be so vital to not only our perceptions, but memories as well.
Re: Abstraction and Hearsay
ReplyDeleteI found this section particularly informative in its explanation of how concepts (i.e. “good”, “truth”, “beauty”) that cannot be grounded on the basis of direct sensorimotor invariants, become categorized through “[tracing] the chain of abstraction that takes us from categories acquired through direct sensory experience to those acquired through linguistic hearsay”. An issue I had with the previous article on “The Symbol Grounding Problem” was how to ground more abstract concepts, whose understanding transcend direct sensory interaction. Abstracting from a sensory experience through a process of linguistic hearsay offers a viable solution to this problem, while still categorizing things, in the most basic form, through our sensorimotor capacity. From categories that have already been grounded in the sensorimotor, we expand and combine grounded concepts to build more abstract categories. But to what extent is “linguistic hearsay” affected by our environment? Can what our environment or culture affords us cause language to actually limit category learning? The commonly used example is the many words used to categorize varieties of snow and ice in the Inuktitut culture. Ultimately, this goes hand-in-hand with the Whorfian hypothesis, but I am left confused with how it would be resolved.
Regarding: “Although basic color CP is inborn rather than a result of learning, it still meets our definition of categorization because the real-time trial-and-error process that "shaped" CP through error-corrective feedback from adaptive consequences was Darwinian evolution.”
ReplyDeleteI believe this paper is right about basic color CP, still, I am curious to know more about to what degree color CP is inborn. Is the inborn basic color CP regarding the inborn ability that we see the visible spectrum 400-700nm as colors and we perceive them as a continuum (but with distinct colors) from red to violet? Isn’t the basic color categorical perception (the ability to divide the visible spectrum into colors like red, orange, yellow, green, blue…etc) learned?
I would like to share an interesting story about my father. My father has red-green color deficiency (not blindness). He and I are very curious about his “weird ability” to see colors in a different way than the rest of us. When I was small, I kept showing him red or green colored pictures. I was expecting to see if he is going to say that is a blue or yellow picture (so that I can laugh at him). Then, he said: “of course I know it’s red (or green)!”. He is good with the very saturated red or green (The very saturated red and green color we learned in kindergarten) but just bad with the red/green when the context is dimmed, or if the red-green colors are more brownish/less distinguishable. But then it brought me to the question whether he is really seeing what we see as the very saturated red. I can never know whether he really sees the red color. It could be the case that whenever a very saturated red is showed to him, what he sees is some brownish color (with a particular frequency) that he recognizes, and after being laughed by people who tease his color perception for years, he learned that this brownish color is called red. If so, then wouldn’t his basic color CP becomes a product of learning?
We all feel natural to divide rainbows into colors when we see it, even when we were infants who haven’t learned how to name those colors. I agree that color perception is inborn, but how do we consider the basic color CP (the ability to decide whether this light frequency belongs to red) in people who have color deficiencies? Instead of saying them as exceptional cases, I have a feeling that they might be giving hints to us that basic color CP could be an ability that needs a lot of learning, a lot of trial-and-error to shape, and the set of cut-offs that divides the light frequency spectrum into separate colors is a product of learning.
(On the contrary, it could also be possible that my father has the inborn basic color CP, the set of cut-offs to divide colors, which he later realized that his set of colors is so wrong and he switched to the new sets of color CP that divides browns into red and green so as to follow the norm, the most common color set that we have. But then, how can we know which of these cases is true?)
Alison, we learn the names of colors, but we aleady see them as qualitative bands in the rainbow. The feature-detectors are innate. We don't need to learn them.
DeleteIf your father has a color perception deficiency, then there are some colors we can distinguish that he cannot. What they look like to him, we cannot know (but we cannot even know that for sure with people who have normal color vision like us!).
Someone who has a red/green color discrimination problem may be seeing both reds and greens as red, or as green, or as some other quality. They will still be able to say that a rose is red and not green because they have learned that "red" is the color of roses. (But they could be tricked by a green rose.) But they do not learn to see the colors we can see; and we do not learn to see them either.
What I am drawing out of this article that I find specifically interesting is the apparent levels of categorization and abstraction that can lend itself to what we consider our consciousness. It seems categorizing any prototypical object allows inductive inferences to be made about that object and its prototypical features. Would it be fair to look at human consciousness that seems to arise out of categorization as a hierarchical system where the most base categories allow us to draw the most salient information about something's features and over time and development we work our way up to conceptual and vaguer 'difficult' concepts like beauty/justice/ethics etc etc. where the inferences are incredily generalized and perhaps less related to actual physical features. The issue with my understanding is that if this is the case, where is the line drawn for other species we argue are not conscious? ie. If a dog is capable of recognizing food or a ball, another dog, good/bad in terms of displays of fear; how is this not a base form of categorization and at what level in the hierarchy does the distinction exist? Are we merely conscious because we are capable of distinguishing concepts that in themselves have no physical features that we can assign to them?
ReplyDelete
Delete@Manon. What you’re asking about is where we draw the line between our own ability to categorize vs. a dog’s ability to categorize. Animals can have categories, as they learn how to react to a physical stimulus, and are able to discriminate between stimuli. However, because they do not have language to ground the symbols they are reacting to, they can’t have categorization in the sense that we think of it as learning, cognition, and abstraction. To sense something doesn’t meaning it is grounded. To have a referent doesn’t mean it’s grounded. A dog can recognize the difference between food that it can eat and an object that it cannot eat, and it could even recognize someone saying “food” meaning there being food there. Yet, it is unable to abstract certain features of the food and apply it in other areas. A child might eat mushed carrots in one context and boiled carrots in another, recognizing them as variations of what we imagine as a carrot. Whereas a dog would not be able to abstract his or her food as belonging to this or that category.
I don’t think we’re arguing against how ‘conscious’ another species is, but rather their ability to cognize and categorize in the way that we do, i.e. using language.
From the paper and class discussion, we know that categorization can influence similarity, and it seems so that similarity can also influence categorization. However, the two differ: as in that categorization is all or none, whereas, similarity is relative judgment and is continuous. In definition, we also know that categorization is an absolute discrimination and just noticeable difference (JND) is relative discrimination, which is unlimited and varies widely. In the section where we discussed the possibility to increase our capacity to make categorizations, it seems evident that despite the dimensionality increase, it will have its limit and would never approaches the resolution power of the JND of sensory discrimination.
ReplyDeleteBut, it seems that the paper suggests that we can attempt to increase our capacity in recoding by overlearning bigger chunks. Supervised categorical training allows us to abstracted the shared features, to find that invariants underlying the variations. And if categorical learning is successful, the features are weighted again and “the inputs are recorded, just as they are in the digit string memorization.” This sentence was curious to me, as it brings it back to the chunking example with the binary digits. That because of chunking we were able to enhance our memory by recoding, hence augmenting our capacity to make categorizations. What makes me wonder is the possible affect of overlearning: such that the more we train the learners on an induced categorical perception, will there be a point where memory no longer plays such a crucial role in our capability of categorizing? That it will become eventually autonomous to react heuristically to categorize the stimuli without trying to remember what the weighted features are? (Or perhaps this will be answered by Marie’s experiment?)
I’m curious about how neurological anatomy and processing might play into categorical perception and colours. Is it possible that categories might be defined if their perception uses separate neural processing systems in the brain? I guess then that this would be considered a question about evolved CP. So, for instance, because colour and black and white are shown to be processed separately in the visual system, through a system of parallel processing, might it be true that that separate anatomy is the reason why we can think of black and white as one category and colours as another? And perhaps it is why there is a starker perceived difference between blue and white, whose perception involves completely distinct neurons and anatomical connections, than that of blue and green, whose perception involves the same neural connections. In this way maybe it makes sense that some categories in cognition come up from evolution, rather than are learned.
ReplyDeleteI guess it depends on what category you’re using right now, so if there are parallel pathways it means that you perceive black and white differently not necessarily in two categories. Like if there’s a list of words of “black, white, round, square”, you would say “black and white” are of the same category because they are both colors whereas the rest is “shape”. So maybe that’s the difference between categorization and categorical perception? Actually I’m a bit confused right now I will call for help @ Stevan Harnad: could you please explain for me the difference between categorization and categorical perception again? I know that categorization is doing the right thing with the right kind of thing and categorical learning is when we learn categorizations (most are not innate) and categorical perception. But is it possible for things to be categorized into 2 categories according to categorical perception but of the same category according to categorization? Like silver and gold are both metals but they are perceived as different things? Thank you very much!!!!
DeleteLucy, color CP could be generated by anatomical differences, or physiological differences, or biochemical differences. With color, which is mostly innate, we know most of the story. The trickier case is learned CP, where something has to change as a result of learning.
DeletePeihong, Categorical perception is a perceptual effect. It occurs if (1) physically equal differences are perceived as being bigger between categories and smaller within categories (separation/compression) (this usually occurs along a continuum, and it is innate) or (2) differences are perceived as being bigger between categories and smaller within categories (separation/compression) when we compare before and after category learning.
17. Abstraction and Amnesia
ReplyDeleteBorges’ story concerns Funes who cannot forget things. The idea is that Funes shouldn’t have been able to speak because our words pick out categories on abstract bases, hence people without this ability are able to speak because they are not caught up in the cognitive over processing implied in this example. Though, these categories are also learned, but people apparently have an innate capacity to learn these categories.
Let’s say a person 1 had a learning disability and another person 2 who suffered a seizure and sustained brain damage. As a result, they can see in colour but don’t know what the colour is. They can see the entire spectrum, but cannot name the colour. Given that they cannot categorize colour but can identify that, say, and apple is red, does this theory mean that they would be able to learn the categories (again) in both cases? Would these categories come back due to our innate capacity to identify category boundaries?
Marcus, I think this is a very interesting question. By recognizing the differences between the colors, or perceiving the differences between the wavelengths of the spectrum, the individuals are both categorizing. The brain damaged individual although not able to "name the color", could probably still point out objects with same/similar colors. The deficit in the ability to name, or assign a language symbol to them, could just reflect damage to the language producing area of the brain.
DeleteRE: Learned Categorical Perception and the Whorf Hypothesis
ReplyDeleteSo I'm very confused at this because if we can create the boundaries of colors based on culture or language, does that mean that everything in reality is like this? If something is hard vs something soft, are those influenced culturally too?
This hypothesis really enforces how the world is really just what our sensory system makes of it. But what confuses me is that some things are concrete realities. Like if something is gold and something is silver, and a culture doesn't distinguish the two, then they are in the same category. But from an actual scientific approach, they are. Where is the line between human perception and actual hard science?
I think this is kind of related to (19) feature selection and weighting from the article. Our sensory systems utilize or place a higher importance on particular features when making categorization judgements, but there are many things in the world which are really there but whose features we totally miss (first thing jumping to mind is microorganisms – they’re so small our feature detectors can’t even pick them up, but that doesn’t mean they aren’t there). So although some features are not particularly salient to humans, doesn’t mean that those features aren’t true invariants in the world; it just means that they aren’t being used to categorize. These salient features allow us to make categories based on our perception/sensation, like soft vs hard or light vs dark. The Whorf Hypothesis is making claims about how we interact the world and the affordances/sensations available to us, but this does not actually take away from or change all of the invariants associated with an existing object. I think it is just about how our culture and language may be involved in the salience or weighting of certain features.
DeleteIf cognition is categorization, then this seems to be another reason to focus on T3 rather than T2. The sensorimotor capacities that are present in T3 are necessary to make sense of our environment. Although it seems to us like an easy task, at any point in time we are interpreting an abundance of information from all of our senses at once. For example, when you’re standing on a street corner, you’re hearing a bunch of different sounds at one moment, but it still seems easy to distinguish where each sound is coming from. In some cases, information from one sensory modality might override another. A classic example of this would be the McGurk effect. The McGurk effect demonstrates that what you hear can be influenced by what you see. Let’s say you are standing in front of a person who is moving their mouth by tapping their lips together, and you hear the sound “Ba”. Now, if you take that exact same sound, but play it over a video of someone opening their mouth and biting their bottom lip, you get the illusion that you are hearing the sound “Va” even though you are listening to the recording of the sound “Ba”. This helps us to understand what happens when our senses are providing us with conflicting information; information from one sense ends up dominating. Therefore, the ability to experience more than one modality at a time is beneficial to categorization. Seeing as this sensorimotor ability is afforded in T3, and seeing as cognition is categorization, then it seems logical to focus on T3 when it comes to reverse engineering cognition.
ReplyDeleteI’m also curious to know whether the example of the McGurk effect could provide support for the motor theory of speech perception.
The efficiency yielded by our innate ability to categorise (and therefore, cognise, as argued by Harnad) is, I think, undeniable – once we have defined an archetype of an object (and subsequently its sub-objects, as we continue to categorise further and further down), it becomes more efficient for us to recognise things, and to some extent, concepts. This instinct to categorise in order to save time has caused some trouble for us humans throughout history within the social ranks – today we champion individuality, and sharing anecdotes and life stories of minority folks to remind others that they, too, are full and complex human beings. It is all too common, within minority groups, to categorise other human beings into groups and simply stop at that without getting to know the individual. This sort of categorise-and-stop process, if taken too far and with negative connotations, is exactly the type of behaviour that leads to harmful prejudice; when it is so easy (literally easier, and more efficient, though in most cases incorrect) to think of a certain group as "all the same".
ReplyDeleteSlightly depressing social justice musings aside, it interests me that the extent of categorisation varies from person to person. In studies of prosopagnosia, researchers reach various conclusions about a prosopagnosic patient's discriminatory ability "within-category". It's interesting that, when someone learns more about a particular type of object (say, vintage cars), such a learning process inevitably involves dividing up the formerly generic "vintage car" category into smaller and smaller subcategories. An individual who knows the intricacies of all the different categories of vintage cars and the way these categories interact is someone touted as an expert: indeed, perhaps we could consider any expert as just a person who has categorised a certain genre of object or concept to an extraordinarily high degree.
“But, at bottom, all of our categories consist in ways we behave differently toward different kinds of things, whether it be the things we do or don't eat, mate with, or flee from, or the things that we describe, through our language, as prime numbers, affordances, absolute discriminables, or truths. And isn't that all that cognition is for -- and about?
ReplyDeleteHow do we account for the amount of human knowledge, thinking and “cognizing” about things such as abstract concepts? These things have no direct use to us, and arguably cannot be sensed in any way, or described using senses. How can we reconcile this idea with cognizing about things like theoretical mathematics or philosophy? Where does our knowledge of categories involving abstract concepts come from, since it doesn’t seem like it could be sensory experience (at least not with the concepts themselves)? It seems like there are many ‘things’ we cognize about but do not sense, interact with, or directly experience – How do we explain these?
The article enlightened me of the extent to which categorization plays on cognition. If cognition is solely categorization then would a type of sorting machine be able to pass T3? Given the said sorting machine has sensorimotor capabilities, and has an algorithm promoting learning by feedback from the most salient features (possibly predetermined like vision is for us). Also as a side note we can still categorize something as red even if we are not physically seeing the red. Some photographer took a photo of strawberries and removed all red pigment from the photo. One version is in blue and green tones, the other greyscale and in both the colour of the strawberries is distinctly red. This is just another fun trick the mind plays on us but it shows the importance of categorizing things such as the categorical categorization of colour (given our knowledge about things like shade , sun and observation angle). The ugly ducking argument really illustrates the need for a sorting machine to be able to ignore most features or focus only on the relevant ones.
ReplyDeleteFrom this article, it seems like categorization and CP largely rely on what our sensorimotor system affords and way it allows us to interact with the world. The principle role of the sensorimotor system extends to cognition if we see categorization as the main activity of cognition. This makes me think that any question surrounding whether robots or artificial systems can cognize like humans, relies on our ability to program them with capacities that reflect the functioning of humans’ sensorimotor systems. If a robot is built with some “visual” capacity, but not with the same limitations on the system we have, it should be expected that what the robot can identify, and thus categorize, would be different. So, maybe in experiments/ observations where it appears a robot can’t correctly identify and categorize, the robot may in fact be categorizing, just in a different way which is afforded by their system. If we wanted to build a robot that cognizes in the way we do, we would need to build its capacities to have the same limitations and afford the same sensorimotor experiences as humans. Differences could have the power to derail categorization ability, and ways of categorizing that stem from significantly different sensorimotor systems may be unimaginable to humans.
ReplyDeleteOn the topic of bilingualism and categorization, do you think it’s possible that bilinguals could have two different ‘systems’ going on at the same time, giving different output to the same input? I read about a study where bilingual speakers gave different questionnaire answers, with the initial one in japanese, and then 6 months later in english, about how they’d complete different sentences. The sentences remained the same, but their answers varied depending on the language they used. Some answers were more family-centric (Japanese), and others were more self-centered (English). I’m curious about what could happen when, for example, a person is speaking French, but in an environment where they usually employ English. Could “French behaviour” and mannerisms still be accessed, even though the person is thinking and actively using French? Here’s the study in question- https://www.psychologytoday.com/blog/life-bilingual/201111/change-language-change-personality
ReplyDeleteQuestion: I am confused about the distinction between supervised and unsupervised learning. Is a category unsupervised because it was previously supervised, but became unsupervised with enough exposure and repetition? Is unsupervised categorization the same as innate categorization because of predetermined sensorimotor affordances?
ReplyDeleteThis reading was extremely clarifying for me. “Some features are selectively enhanced, others suppressed, thereby bringing out the commonalities underlying categories or kinds.” I found this to be an excellent explanation of how categorization works. Supplementing this with the examples of the character by Borges and the person observed by Luria perfectly demonstrated the importance of categorization in our day-to-day lives, and nicely led up to the finishing sentences that categorization is essentially what cognition is for, why it exists. Explaining Luria’s S. And Borges Funes as ‘handicapped’ rather than having an advantage in memory made perfect sense in the context of Harned’s explanation of categorization and cognition, and really enriched my understanding of the importance of categorization.
ReplyDeleteThe discussions about learning and supervised and unsupervised categorization made me think about how they their relationship to computation.
ReplyDelete“The unsupervised models are generally designed on the assumption that the input “affordances” are already quite salient, so that the right categorization mechanism will be able to pick them up on the basis of the shape of the input from repeated exposure and internal analysis alone, with no need of any external error correcting feedback.”
By this definition of unsupervised categorization can a computer be a categorization mechanism? It seems to me that this definition is compatible with the features of a computer: In this model, the computer would pick up the affordance of the stimulus based on its shape which is consistent with computation being shape-based and the subsequent internal analysis mentioned above could be the rule-based operations and changes of status done in a Turing machine. The troubling part, however, is about the affordances of the stimuli for a computer which does not have any sensory motor capacities for receiving an input or executing an output. Is it the case that inputs do not have any affordances for a computer due to the lack of sensory motor capacities or is it just that their affordances are different from what they afford for us as sensory-motor beings.