Harnad, S. (2017) To Cognize is to Categorize: Cognition is Categorization, in Lefebvre, C. and Cohen, H., Eds. Handbook of Categorization (2nd ed.). 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.
See also: Jorge Luis Borges Funes the memorious
Categorical perception is proposed to be constructed through the simultaneous enhancement of invariant features and the reduction of salience of non-invariant features. This allows us to focus more attention on the common features between different objects of the same kind, making categorization possible. This capacity of abstraction is proposed to be enabled by the mechanisms of supervised and unsupervised learning. We realize this capacity’s importance through Borge’s example of “Funes”, a person who cannot abstract and is therefore considered to have an infinite rote memory, condemned to store every successive instant of his experience forever.
ReplyDeleteWhile I think I have understood that what allows us to focus on the right invariant features for categorization are our sensorimotor interactions, I am unsure of the relation between this concept and Gibson’s “affordances”. From what I understand, equating invariant features and “affordances” would mean that we unite objects together, making them part of a category, on the basis of the sensorimotor interactions that they allow us to have/afford us. This makes sense with our previous discussions in class, but if anyone can further clarify these concepts it would be greatly appreciated!
Mathilda, invariant features are features that distinguish members from non-members of a category, or distinguish members of a category from members of other categories.
DeleteUnsupervised and Supervised (or Reinforcement) Learning
Unsupervised learning alone – repeated passive exposure -- can only detect and highlight which features are more frequent, and which features are correlated with one another in the sample to which it is exposed.
For example (and I’m deliberately using trivially obvious categories and features to give you the idea), there are red, green, and yellow apples, but more red apples than green and yellow ones. And brown apples tend to be mushy and taste bad. And no apples are blue or crescent-shaped, whereas bananas are crescent-shaped and first green, then yellow etc.
Unsupervised learning alone can selectively detect all these frequencies and correlations from being exposed to them over and over, but it cannot detect what’s an apple and what’s a banana. That’s a correlation too, but not a correlation between a feature and another feature but a correlation between features like yellow and crescent-shaped and what is the correct thing DO with something that has those features (e.g., eat it).
To detect a correlation between features and what you need to DO with something that has those features, you need supervised (or reinforcement) learning, which is no longer just passive exposure. There is corrective feedback from the consequences of doing the right or wrong thing with the object that has those features.
You already know that a way to learn the correlation between a stimulus and the right response to it is called “reinforcement learning.” In reinforcement learning, repeated trial and error increases the tendency to make the right response (which is rewarded) and decreases the tendency to make the wrong response (which is not rewarded). The corrective signal for a right and wrong response is the reinforcement.
Supervised learning is a causal model of a system that can actually do this. (It’s called reinforcement learning when it is done one trial after the other in real time, but it can be done by a computer in a batch, by feeding it a huge number of “labelled” input-output pairs, in which the correct response (eat, don’t eat) is already paired with each stimulus in advance, so that the correlations between the input’s features and the right response can be quickly learned. (I have left out some technical detail on what the mechanism is that strengthens the connection between the features and the right response. It is called “error back-propagation.” I’ll tell you a little more about that if it comes up, but you can already look it up if you’re curious.)
Gibsonian “Affordances”
DeleteIn a sense there is no such thing as pure unsupervised learning. It would be possible only if, like the passive kitten in Held & Hein’s (cruel) experiment, you were the one being transported without moving your body (or even your eyes) and images were passing over your retina without you doing anything. In the real world (without heartless experimenters sewing kittens’ eyes shut), there is always a sensorimotor interaction with any sensory input. Eyes make saccadic and pursuit eye-movements, scanning the input, the body moves, the hands manipulate things. This sensorimotor interaction is not just sensory; nor are its features.
The features of a sensorimotor interaction were dubbed “affordances” by (James) Gibson. They are neither purely sensory nor purely motor features. They are the features of the interaction of our bodies with objects outside our bodies. Gibson’s famous example is a chair. What are the “features” of a chair? Four legs (usually), wooden (often), etc. But also “sittability-upon” -- by organisms approximately the size and shape of humans. That is not a feature these same objects afford to elephants or whales. Nor to bees or ants. And pitch dark, bats can perceive the distance, size and shape of objects from the echoes that bounce back from the sonar (sounds) that bats aim at them. This is not something that objects afford to humans at night.
So there are not only sensory and motor features, but also sensorimotor “affordances” that are based not only on the shape of the object but the “shape” of the perceiver.
(Now a subject for reflection: Aren’t all “features” of objects, even the “sensory” ones, actually affordances? And isn’t categorization – doing the right thing with the right kind of thing – all based on affordances?)
"And isn’t categorization – doing the right thing with the right kind of thing – all based on affordances?)"
DeleteCategorization cannot be based entirely on affordances. Features' affordance is determined by "what can be extracted from their motor interactions with their sensory input (to our proper invariance-detector)" (Harnad). Doing the right thing with the right kind of thing can't be based on affordances alone because, most of the time, there are different "correct" ways of clustering the same sensory shadows. If I understand it correctly, this is related to a category's "extension" and "intension"; the former implies that there is an infinite amount of things that are members of a category, and the latter means that there's also an endless amount of features that make a thing member of a category rather than another. And to correctly sort out an ocean of features depends on the demands of the context/situation, not just affordances alone. You mentioned in your reply that "there is no such thing as pure unsupervised learning," the idea that the acquirement of affordance is largely an unsupervised process (affordance = the interaction of our bodies with objects outside our bodies) suggests that affordances alone do not offer us any corrective feedback from trial and error. Therefore, to learn to categorize things correctly, supervised learning is a necessary condition because we always do categorizations "within a certain context" where trial and error feedback is required.
Just wanted to clarify what I meant by "within a certain context.":
DeleteThe reading 6a says that categorization encompasses both absolute judgment and relative judgment. The former is based on identifying an object in isolation; the latter is based on the alternatives amongst which the isolated object needs to be sorted. The alternatives here refer to the context. For instance, a folding chair, when compared to other household items in an apartment, is categorized as a piece of furniture. But when placed among items such as steel ladders and guitars in a WWE fight pit, a folding chair is categorized as a weapon. To correctly categorize a folding chair in the two abovementioned contexts can't be relied on its affordance alone.
Yucen, good points! But “affordances” – what the shape of an object will allow an organism with a certain shape to DO with it – is a supervised, not an unsupervised matter. There are consequences of doing the wrong thing with an object. If you eat something that has the shape of a poisonous mushroom, the “shape” of your metabolism will make you sick, as surely as if you try to open a lock with a key that does not fit.
DeleteYou are right that the features that are enough to distinguish members and non-members of a category are not unique. Categorization is both approximate and underdetermined. Not only could the features you have been using successfully till today turn out not to be enough to resolve the uncertainty in a bigger context of alternatives tomorrow (so you have to update them), but maybe there are different sets of distinguishing features that resolve the uncertainty equally well.
(“Affordance” has become a bit of a Gibsonian cult-word, maybe even a weasel-word, so don’t put too much weight on it.)
The “extension” of a category is all of its members. The “intension” of a category is the features the distinguish its members from its non-members. (Don’t mix up “intension” with “intention.”)
There are two relevant senses of “context” in category learning. The “context of acquisition” of a category is the set of confusable alternatives in which the learner has to learn, through supervised trial and error, a set of features that will reliably distinguish the members from the non-members. The “context of application” of a category is how you use the category in everyday life, once you’ve learned it (and its features). Your WWE example was from the context of application of categories you already have, not the new ones you have to learn, by trial and error.
Amélie, yes, to the extent that the notion of “affordances” makes sense, any “feature” that “affords” doing the right thing with the right kind of thing is an affordance. And they may differ from species to species. The shape of faraway objects in the dark does not afford localization to rats but it does to bats (because of the sonar they can produce). Things that are toxic to us afford nutrition to a buzzard.
"But 'affordances' – what the shape of an object will allow an organism with a certain shape to DO with it – is a supervised, not an unsupervised matter. There are consequences of doing the wrong thing with an object."
DeleteThis made me think of Don Norman's book "The Design of Everyday Things", where he describes affordances in design as "perceivable action possibilities." Norman describes that as a designer, one must make obvious what a button/interface detail "affords" the user, aka make intuitive. If an interface detail does not alert the user to what can be done with it, Norman argues that the system is poorly designed for use. This is particularly important to understand, as even if the "affordance" fits to the users' capabilities, the inability for the user to understand how to use an interface button/slider/toggle/etc. due to a lack of connection to past instances / previously established categories can result in user frustration and thus disengagement from future use.
I took a cognitive science class last semester in which we talked about categorization a lot but I never fundamentally understood why it was all that related to cognitive science. This piece clarified that for me. Categorizing is the process of abstracting certain features in lieu of others that can be glossed over, and relies fundamentally on knowing what to do with what kind of thing–a crucial capacity necessary for survival. Categorizing has always been synonymous to classifying in my head, and I remember learning about a machine learning method of classifying based on the k-nearest neighbors, allowing them to classify an item by a plurality vote of its neighbors. This resonates with the color category section of this piece in which all things with category boundaries are on a spectrum and has a range of similarity with the qualitative properties of one side the neutral point. I wonder if this is really what we do in our brain or whether this is just a computational model. And what does it mean in terms of evolution that the midpoint between black and white is not innate but that the one in between green and blue or hot and cold is? And if this task is already complex enough for a two-alternative comparison, no wonder it gets hard to categorize with the myriads of categories that are available to us!
ReplyDeleteTess, “classifying” is one (of many) forms of categorization. But only if it’s supervised classification, i.e., there is a right and wrong, with positive or negative consequences from classifying correctly or incorrectly.
DeleteAs reading 6a points out about classifying playing-cards, they have lots of features, and you can sort them according to many different features. But remember that the game of life (categorization) is doing the right thing with the right kind of thing, not just sorting (or doing) things as the spirit moves you. Doing the right or wrong thing has consequences.
And remember what “information” means: reducing the uncertainty about what to DO, when doing the right thing matters in some way, e.g., getting a vegan lunch from the many-windowed sandwich machine.
No computational model for category-learning has been shown to scale up to TT-scale yet, so, like all models (and all categories!) category-learning models are approximate (and in cogsci, they still have a long way to go!).
But we already know that models for color learning or phoneme learning (especially vowels) are not representative of most of the categorization tasks humans face, because most of the things that we need to learn to recognize and do the right thing with are not points along a feature continuum (let alone innate pints!). Things have many different features, and we have to learn to detect which features distinguish what things we need to do THIS with and what things we need to do THAT with.
And all those tasks call for learning. The basic colors are innate, but most other categories have to be learned. (We’ll talk about the “laziness” of evolution and the power of learning -- greater even than the power of language, which is itself a powerful form of learning – in Week 7.)
Hi Tess! I had a similar misunderstanding to you about how categorization relates to cognitive science. Categorization is not just about assigning things to a group, but rather knowing what to do once they have been assigned. This quote from the reading helps explain this idea, “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.”
DeleteMy question relates to Prof. Harnad’s explanation of categorization, describing categorization as “doing the right thing with the right kind of thing, not just sorting (or doing) things as the spirit moves you. Doing the right or wrong thing has consequences.”
I understand that this explanation is applicable in some contexts in which the consequences of an action are clear, but how does this relate to a situation in which an action leads to a neutral outcome? For example, how is it more “right” to choose one sandwich over the other?
Kimberley, to categorize is to do the right thing with the right KIND of thing. When doing the right thing matters, and doing the wrong thing has consequences.
DeleteRemember the vegan sandwich machine with the 6 opaque windows when we talked about what information and uncertainty are? If someone you trust knows tells you that the only sandwich is behind an even-numbered window, the right thing to do is to pick an even-numbered window, not an odd-numbered one. And that’s the same as someone telling you that the striped mushrooms are safe to eat.
But if the choice is between 6 different delicious vegan sandwiches you can see, and you choose by whatever you happen to feel like eating at the moment, that is just an ad-lib choice “as the spirit moves you.” Nothing at stake; no right or wrong; no risk; no uncertainty.
In the choice between 6 different delicious vegan sandwiches, can't we talk about discrimination coming into play rather than categorization? As I look at the 6 delicious vegan sandwiches, I might want to pick the largest, and look for "just-noticeable-differences" relating to the relative sizes of the 6 delicious vegan sandwiches, without ever having to categorize, that is make a choice between delicious vegan sandwich of size x and delicious vegan sandwich of size y. (Funny enough fast food does provide us with size categories small, medium, and large, presumably so that we do not do the wrong thing with the wrong kind of thing, for example fail to satiate ourselves, or overeat).
DeleteMy understanding of categorical perception is that it involves both categorization and discrimination. Discrimination is relative whereas categorization is absolute. Discrimination becomes categorical if we can do something right or wrong. It has to be right or wrong to be categorized; it can’t just be not good enough. There are also other things we can do that are not categorizing. For example, continuous motor skills are not categorical but continuous (like dancing, walking, and sprinting). And the ability to copy, imitate, compare and make similarity judgments on things is not categorica. To categorize is to do the right thing with the right kind of thing.
ReplyDeleteMelis, discrimination (perceiving how different two things are) is a matter of degree. It never becomes categorical. Categorizing is identifying one single thing alone, by doing the right thing with it (e.g., saying its name.) See above, about the way learning to categorize can change discrimination.
DeleteI had a similar understanding as Melis when it came to discrimination being involved in categorization, I saw it as more of something that comes with categorization but I now understand that discrimination is something that can be changed by categorization. Adding on to this comment, categorical perception only applies when the matter is supervised, involving a positive or negative consequence, relating to your vegan sandwich example, as opposed to having a situation with no real consequences (i.e. a series of good vegan sandwiches)
DeleteAriane, there is only categorization and category learning, where there is a right and a wrong thing to do. Where there is no category learning, there is not categorical perception. How is this related to information and to supervision/reinforcement?
DeleteAnd what is the difference between category learning and categorical perception?
This relates to supervised/unsupervised learning because supervision is required to make correlations, therefore it is required to make discriminations. In your earlier red/green/yellow apple vs bananas example, unsupervised learning can detect the correlation between a feature and another feature, but not the correlation between features. Supervised learning is required to discriminate the correlation between features (and also what needs to be done with what has those features)
DeleteAlthough animals can surely categorize, language makes categorizing much easier and faster because the information is conveyed more quickly. I believe that this evolutionary advantage is one of many that have permitted humans to adapt so well to many situations.
ReplyDeleteI also reflected on that and how it basically allows for an additional way of learning, verbal instruction. It's pretty interesting considering we use verbal commands to teach pets and the substance of the words means nth and any action could essentially be elicited by another form of sensory input for them.
DeleteTo add on what Melis and Hassanatou wrote about the importance of language in categorization, language as discussed in the article is the reason as to why we were able to do abstraction and hearsay. Without language we would have never been able to categorize things such as “goodness”, “truth” and “beauty” that were categories acquired through linguistic “hearsay”. On top of that, I believe that language also has a big influence on how we cognize categories and therefore has played a big role on how we learn such.
DeleteVitoria what is “hearsay” and how does it make it possible to categorize “truth”? And how does language influence how we “cognize categories” and how we learn. Please read the other replies to get a clear idea of this.
DeleteIn addition to clarifying the concepts covered in class last Friday, this reading was insightful by pointing to a recurring notion in this class: "it is easier to say what a system does [...] than to say how it does it". In this context, the phrase referred to the learning process, which is easier to explain in terms of 'what' than 'how.' However, I found it very fitting regarding the last 5 weeks of class and the subject of cognition as a whole, which attempts to explain 'how' and 'why' organisms can do all that they can do, rather than simply desribing 'what' they do.
ReplyDeleteAmélie, actually, there is at least one candidate “how” explanation for unsupervised and supervised learning: deep learning, with error-correction by backpropagation. Does anyone know what that is?
Deletehttps://www.youtube.com/watch?v=Suevq-kZdIw
https://www.youtube.com/watch?v=yRUUDJfDarU
From what I remember from a class last year (and the youtube videos you linked):
DeleteDeep multilayer neural networks seem to be a possible answer to the “how?” question (though we still do not understand everything). The single-layer perceptron was a start, but once multilayer neural networks were attempted, the computer could categorize successfully through supervised learning algorithms (for example, recognizing and categorizing written numbers). Backpropagation, the systems’ fine-tuning of connection strengths, is essential in this system, allowing the neural networks to approximate any function (theoretically).
This approximates how our brains supposedly use systems of neurons to learn categorization.
Kayla, neural nets have been shown to be able to learn (some) categories by detecting and abstracting their distinguishing features through unsupervised and supervised learning. But what about verbal learning? Time to start thinking about that too.
Delete"To abstract is to single out some subset
ReplyDeleteof the sensory input, and ignore the rest." My understanding of this is that it wouldn't be incorrect to say that symbol grounding requires more abstraction as we move up in symbolism. Like in the example of the "horse + stripes = zebra", there are certain features we ignore and some that we weigh greater in order to form a zebra from those two symbols. This makes sense as this is essentially what categorization is and being able to solely focus on key features is what allows the ability to do “the right thing” respective to each thing.
The idea that everything would basically be the same if we didn’t pull out specific features was kinda trippy. Not a huge interest/concern of mine but could much be said about categorization in/and topology?
Hassanatou, Kid-sib could not understand your last sentence.
DeleteThis reading provides a clear explanation of how we learn categorization. There are different theories of how me might be learning to categorize objects. One of the views proposes that categorization is an innate mechanism that we are born knowing, similar to Chomsky’s argument about universal grammar. Another idea is that we learn categorization by trial and error, guided by feedback through supervised learning. Language, in the other hand, is special since it can also be learned through “hearsay”. While we categorize, we use the method of “abstraction”, where we abstract the invariant features of an object with our sensory motor system to be able to discriminate them and place them in categories.
ReplyDeleteFrom what I understood, none of these explanations on their own are enough to explain how we categorize i.e. it can't be only one way all the way down, therefore we actually use a mixture of all the methods, depending on the situation.
Of all the counterexamples to categories being unlearnt (innate), my favourite is the peek-a-boo unicorn. This idea ties into our concepts from last week because it shows how powerful grounding is. Once you’ve defined and grounded enough concepts (i.e., “horse” and “horn”) you can ground others as well (such as a horse with a horn—a unicorn). Unicorns don’t exist but even if they did, the peek-a-boo unicorn vanishes without a trace once someone tries to perceive it. Yet, the peek-a-boo unicorn is still a defined category. One must hear about it indirectly (“hearsay”) because sensorimotor interactions cannot happen, but this does not impede the understandability of the peek-a-boo unicorn as a well-defined category, showing the power of grounding, the power of language, and the mechanisms that allow us to categorize.
DeleteAlara, what is the differences between innate categories and the innate capacity to learn categories?
DeleteLearning categories through language by “hearsay” is not the same as learning language by hearsay. What is the difference?
How is verbal learning (“hearsay,” instruction) grounded in unsupervised and unsupervised learning (induction)?
Teegan, you got it. But notice that the “Peekaboo Unicorn” does have sensorimotor grounding, even though we never had, nor could have sensorimotor contact with it. How?
Because we could still describe it with words that are grounded? Draw a picture of it, etc.?
DeleteTeegan, a “PBU” is just as grounded to its referent through verbal grounding as “zebra.” That’s the power of (grounded) propositions.
DeleteI was a bit unsure when I tried to answer the questions above. I believe learning categories through language by "hearsay" is associating words to their referents and assigning them labels, thus learning to distinguish the real-world objects mentioned in the "hearsay ." For example, we don't have to taste to know that the green banana is raw and the yellow one is ripe. We only need to hear it associate "green banana" with "nonedible." I'm not quite sure about learning language by "hearsay." I think it is different because the new linguistic knowledge we gain through hearsay will only be verbally grounded in the more primitive symbols of the language. For example, we learn the word "anguish" when we hear someone is going through horrible things. "Anguish" to us is known as a word that's used to describe misery, and there is no categorization involved.
Delete1. Learning categories through language by "hearsay" is the same thing as verbal learning/grounding. This means that referents are grounded through the connections / rules we are told verbally through language.
Delete2. Learning language by hearsay from my understanding is one type of supervised learning. We can be told through error-corrective feedback what is right and wrong (grammatically) by other people. However, learning language is not all based on corrective feedback (as the behaviorists got wrong and Chomsky later argued against with UG).
ReplyDeleteI found the concept of indeterminate's really interesting and I was wondering if indeterminate classifications can be a mechanism to learn new categories? A simplified example is given with the classification of black and white. These colours exist on a gradient and systems are able to classify them as either nearer white, or nearer black. But what about the point in the exact middle, exactly between black and white? This area is indeterminate and cannot fit into either category as black or white. In this case would it be correct to say that a new category would be formed? Neither black or white, but “grey” which exists directly in the middle as it cannot fit into either end of the spectrum. Once this new classification is made, it could form a new category where the system can learn what to do with "grey" as opposed to what it does with "black" and "white". (I apologize if the colour example is a little confusing/simplistic, struggling to think of a better one!)
Sophie, categories based on boundaries along a continuum are a dramatic example for the separation/compression effect of CP, but they are not really representative of most categories we learn. Look in a dictionary: All content-words (nouns, verbs, adjectives) are category names. How many do you find that are located along a continuum with boundaries between them, like colors?
DeleteMost categories have many features, and among those are features that distinguish members from non-members. Among animals, for example, mammals are distinguished from non-mammals in that they give birth to live young rather than laying eggs (or splitting).
But, yes, in a world full of things with multiple features where some members of C and D are not distinguishable using the current approximate set of features, the features either need to be updated to resolve the C/D uncertainty or there is need for a 3rd category, E (and the features that distinguish the members of each). [See the preceding reply]
From what I understand, “vanishing intersections” refers to the idea that some categories are mutually exclusive, and have no overlap in features. For example, dogs and cats have many intersecting features (both are animals, both have fur, both can be pets, etc.), while dogs and books have no intersecting features (the intersection has vanished). Fodor applies this idea to categorization by saying that because some categories have vanishing intersections, they cannot be learned through experience. They cannot be used in reference to each other and do not provide feedback to each other because they have no overlap. Therefore, Fodor believes that some categorization abilities must be innate.
ReplyDeleteI am interested in the feature selection and weighting section of the reading. “Our sensorimotor systems do not give equal weight to all features; they do not even detect all features” (Harnad)
ReplyDeleteWatanabe’s “Ugly Duckling Theorem” states that if one objectively made a list of all the feature differences and did not weigh them, then everything is equally (and infinitely) similar to everything else. However, this does not seem to be an interesting point. If everything is equally identical to everything else, then there appears to be no point in categorizing; we would not successfully learn what the right thing to do with a certain thing is. For this reason, it makes sense that our sensory systems do not give equal weight to all features but also when remembering due to adaptive and evolutionary reasons, I assume. Harnad states that salient features are shape and color, whereas less salient features are spatial position. For example, I remember what the stew that made me sick smelled like, but I don’t know whether it had carrots or not.
However, if the number of leaves on my plant allowed me to differentiate between mine and my sister's –would that be a more salient feature than the leaf color? Can we possibly say that salient features depend on the context?
I was wondering about the feature saliency question myself- it seems to clash with the idea that conceptual kinds are dynamic. However, I would say, from my limited understanding, that context-dependency is one of the primary features of category systems (at least via the feature-detection view thereof). So if features constitute categories, we might look for or even view the same feature as a different input.
DeleteThere has been a lot of interesting work on the visual system that shows that we will detect the same shade as lighter or darker depending on context. This seems to back up Dr. Harnad's claim that "categorization occurs when the same output occurs with the same kind of input, rather than the exact same input."
It also seems to take it a bit further, where we could say that the exact same input is not always filtered into the same kind, depending on other perceptual filters that may (or may not) be applied.
Kayla, for a more vivid idea of why features have to be selected and weighted (whether innately or through learning), please read Funes the Memorious.
DeleteJacob, think of both features and context in terms of resolving uncertainty about what to do with what kind of thing. The uncertainty always depends on the context of confusable alternatives – on what can be confused with what. The features that can distinguish the categories have to be detected and abstracted (Funes cannot do that).
And yes, what to do with a thing, including what it’s called, may differ in a different context of alternatives. Different features may have to be detected and abstracted.
Re-reading Funes the Memorious helped put some ideas into perspective. We have to be able to select and weigh features to abstract items. As Funes shows, if my concept of “dog” was tied to when I first saw one (a golden puppy laying down on a mat with a ball in its mouth), the next time I saw a dog (a beagle being walked outside for example) I would not be able to characterize that in the category dog: “selective forgetting, or at least selective ignoring, is what is required in order to recognize and name things.” Weighing and selecting which features are necessary for categorizing is essential for generalization across situations and feature variation.
DeleteThis reading made me think about categorization in a different way than I had in the past. I learned that it involves specifically doing the right thing with the right kind of thing, and based on the thing you do, there will be certain consequences. If you do the wrong thing, there will be a negative consequence in some form or another, providing feedback. Through trial and error, eventually, you will get to the level of abstraction that is required to have the category grounded. It's also interesting that language acts as a sort of shortcut in our learning and categorization such that we don't have to go through the process of sensorimotor learning (trial and error), but we can understand relatively quickly through hearsay (this is likely one of the linguistic factors responsible for the massive success of the survival of humans). However, not everything can be grounded this way; the linguistically-based categories themselves must be grounded in sensorimotor categories.
ReplyDeleteAlexander, good summary.
DeleteBoth positive and negative consequences provide corrective feedback in supervised/reinforcement learning, and help your brain detect and abstract the features that distinguish the categories. Language (hearsay) only works if the words in the verbal definition or description of the new category are themselves already grounded (either directly, i.e., through supervised sensorimotor learning, or indirectly, by verbal description of the distinguishing features -- whose names have already been grounded).
I don't agree with innate categories, in that we were all born with the capacity to respond differentially without ever having to learn to do so. It feels like something to learn and to specifically learn a category and categorize them, and if this was innate, I don't think we would feel this feeling of increasing levels of abstraction and understanding in our minds. I do agree with universal grammar (UG) because of the poverty of the stimulus and the fact that we're hardwired to be experts at language, which is the most important factor that lead to the success of our species.
ReplyDeleteHi Alexander,
DeleteWhile I agree that certainly not all categories are innate, I think that at least some type of innate categories are present. Namely, these are most basic shapes and orientations. Simple cells and complex cells in visual cortices are neurological evidence - they can respond to certain groups of basic shapes, edges, and orientations. A most basic 'accordion effect' of grouping similar patterns and discriminating different shapes loom from their outputs. I think that categorizing these basic building blocks is a requirement to categorizing more complex patterns.
If you agree with Universal Grammar, you're already conceding the point that some categories are inborn, no?
DeleteJust wondering if I'm on the right track here in addressing this point: I think that agreeing with Universal Grammar in the sense where the poverty of stimulus argument is supported by evidence in children's development in their language, yet there is still no clear evidence that this supports the idea that other categories are also innate in this way.
DeleteAlexander, you are right that almost all of our categories (hence their features) are acquired through learning. (Look in a dictionary: for how many of those categories is it believable that we were born with their feature-detectors already built into our brains?) Evolution is lazy(Week 7): It is easier, more flexible, and cheaper to evolve the capacity to learn the features that distinguish the categories we might eventually need than it would be to evolve their feature-detectors already built in genetically.
DeleteHan, you are right that there are built-in feature-detectors in the brain, but distinguish between learning to ground a category -- say, “circle” -- and the feature(s) you use to ground it. The features might include curvature (along with other features of the referent of “circle” that distinguish them from the referents of “oval” and “square,” if those are among the categories in the context of confusable alternative categories from which you must learn what a “circle” is. Curvature may have an innate visual feature-detector in the brain, which can be used to learn and ground the category “circle.” But “curved” is not yet a grounded category! It is just a feature being used to ground the category “circle.”
You can go on to learn the category “curved” too, but there the context of confusable alternative categories is different, and probably includes the category “straight.” The features for distinguishing the category “curved” from the category “straight” may be innate too, but they are not the same as the features for distinguishing “circles” from “ovals,” any more than they are the same as the features for distinguishing “eggs” from “golfballs,” though there too they have some features and feature-detectors (whether innate ones or learned ones) in common.
The thing to bear in mind is that it is always the uncertainty in the context of confusable alternatives in learning a category that determines what the distinguishing features will be. A different context of alternatives may require a different set of distinguishing features. This matters most when learning the category. If it’s an important category, encountered often along with its confusable alternatives, the distinguishing features have to be reliable ones, otherwise you risk doing the wrong thing with the wrong kind of thing.
Gabriel and Sara note that UG (or, rather, sentences that obey the rules of UG) turns out to be an innate category. But there the story is a bit more complicated (Weeks 8 and 9). See “Poverty of the Stimulus” in other replies.
DeleteWhat is UG?
I used to be confused about universal grammar in my early undergrad years, mistaking it for the concept that “all languages have the same grammatical rules”. However, throughout my undergrad, classes such as introduction to speech science class and cognition classes taught me this is not the case. Furthermore, the video below was a video I remember watching to help me.
DeleteIn sum, the video basically explains universal grammar works on the premise that all languages will more or less have identical features. As children, we are able to learn the language taught to us by our parents and environment in a natural way -> we have this innate predisposition of attaining language. From here, arises the language acquisition device, i.e., the innate ability to be prepared to learn the language. In such cases, we have the innate ability to begin learning language from predisposed grammar rules within us which are the “identical features” shared among languages. This does not mean grammar rules that refer to exact semantics (as an example), but rather the more general aspects of language. This includes the general idea of nouns, verbs, adjectives, etc.
https://www.youtube.com/watch?v=reYP-kKRhTk
I think this reading in a sense actually brings us back to the problem about whether cognition is computation (computationalism). We, T3 and T4 are sensorimotor systems (from week 2), and from this week we know that we, T3 and T4 are sensorimotor systems that do the right things with the right kings of things, ie. cognizers. So, cognition is not computation in the sense that a computer can not physically do the things we do. On the other hand, as we have seen from this week reading, I can't help but think learning (more specifically, categorical learning) is computation. For example, the artificial neural networks that can detect features from shapes of objects (unsupervised learning) and trial and error (supervised learning). But isn't all of this just symbol manipulations (not grounded because nets are not sensorimotor systems)? A neural net still doesn't know what means anything even though it could categorize?
ReplyDeleteReflecting upon my own comment I think I might know why I was initially confused. Suppose we have some neural nets and we run supervised and unsupervised learning algorithms with them. Then suppose they were able to categorize objects as the result of categorical learning. But what was demonstrated here is an analogue of our cognitive capacity (to cognize is to categorize). The neural nets can not do what we can do and furthermore, they don't have feelings like humans do either. Nevertheless, what neural nets demonstrated is the mechanism of the How. How do we categorize and hence cognize. I think I was most likely mixed up of the What and the How. Then it makes sense because then neural nets are just useful tools for us to learn insights about cognition processes. It does not really matter it is purely computation, since they were not meant to not Be the same things as real cognizers. That being said, new questions arose...Namely if these networks are not doing what we are doing, then what exactly are they doing? Or maybe it is related to Marr's level of analysis? The neural nets are at the level of computational analysis of cognition. Where we, namely real human beings ('cognition) are at both the level of computation, algorithm and implementation level of analysis. So the nets are doing partly of what we do, but not all.
DeleteCynthia, the correct conclusion from Searle’s CRA is that cognition cannot be all computation – not that cognition cannot be computation at all. So part of the reverse-engineering of cognition could be neural nets – whether implemented as a parallel-distributed network of interconnected nodes transmitting activations, or as a computational simulation of them. Either could be doing feature-detection and filtering for category learning. And it could be inside Kayla’s (T3) head or in the brain (T4) – alongside everything else, notably the sensorimotor system.
DeleteBut T2, even if it could be passed by computation alone (“Stevan Says” it couldn’t, because of the symbol grounding problem), would not be thinking or understanding. And neural nets alone – whether physical or computational – could not pass T2. They are just potential mechanisms for learning the features that distinguish between categories.
3 ways in which we learn to categorize are by trial and error with feedback, exposure, and verbal 'hearsay'. I am curious to break this down to discover how much of learning categorization, such as our ability to abstract in our everyday lives, is from trial-and-error vs direct feedback. Additionally, during development, is there a cut-off or optimal period for which this mode of learning categorization must happen? My curiosity comes from a study discussed in class that described children who are deprived of language and the age at which they would be able to recover and achieve the same levels as children who are not deprived. What role does this inability to adequately learn language play in categorization learning and consequently other aspects of their lives. Would this affect their ability to categorize if they lacked the ability to verbally express their sensorimotor experiences, or to fully comprehend the verbal feedback? Could they attain the same behavioral performance as their peers if exposed to more instances of trial and error (with vs without reinforcement), making up for the lack of categorization learning via verbal feedback?
ReplyDeleteHey Sepand, I was actually thinking about this. The reading had also made me consider what it would mean for those with amnesia. I remember a case where someone couldn't form new memories but overtime was able to improve on the piano if my own memory serves correctly. I wonder if the brain's systems What does it mean that their sensorimotor experiences can be continued and even memorized but the person may not remember.
DeleteSepand, explain unsupervised, supervised and verbal learning to kid-sib.
DeleteVerbal correction of trial and error is still supervised learning, not verbal learning.
Only humans have language, but many species can learn categories: how?
A feral child with no human interaction could still learn categories. How long the critical period is for language-learning, I don’t know. (Maybe someone taking Developmental can answer?)
Friedmann, N., & Rusou, D. (2015). Critical period for first language: the crucial role of language input during the first year of life. Current Opinion in Neurobiology, 35, 27-34.
“The critical period for language acquisition is often explored in the context of second language acquisition. We focus on a crucially different notion of critical period for language, with a crucially different time scale: that of a critical period for first language acquisition. We approach this question by examining the language outcomes of children who missed their critical period for acquiring a first language: children who did not receive the required language input because they grew in isolation or due to hearing impairment and children whose brain has not developed normally because of thiamine deficiency. We find that the acquisition of syntax in a first language has a critical period that ends during the first year of life, and children who missed this window of opportunity later show severe syntactic impairments.”
Helen, “procedural learning” from repetition can still occur with dense amnesia.
To answer "Only humans have language, but many species can learn categories: how?":
DeleteCategorization is crucial for survival because it allows organisms to identify and classify things in their environment so that they know how to interact with them. Categorization has consequences, as an incorrect categorization can lead to incorrect actions that may be detrimental to survival.
There are three methods to learning categorization: unsupervised learning, supervised learning, and verbal learning. Unsupervised learning involves repeated exposure to a category to detect correlated features. Supervised learning involves receiving feedback (consequences) from making a classification. Verbal learning involves being explained the features that distinguish a category. Verbal learning is the fastest way of learning categories, but it requires the use of language, which is exclusive to humans.
While only humans have language, other species are still capable of learning categories through supervised learning and unsupervised learning.
Sophearah, Correct.
DeleteThe biggest confusion I’m left with upon reading 6a is the colour example that is referred to a few times throughout the reading. At first, I understood that because the difference between black and white is not as abrupt as is the difference between blue and green, we can experience categorical perception (CP). But I don’t quite understand how this relates to the later explanation of recoding and feature selection… I understood recoding as the process of using supervised learning (ie. trial and error) to re-weigh the features with which we categorize, but how does this apply to the colour example? Perhaps my understanding of recoding is incorrect, so any clarification would be greatly appreciated!
ReplyDeleteFrom my understanding in the reading, our sensorimotor systems do not give equal weight to all features, and they do not detect all of them either. And, of the features they do detect, some stand out more than others, i.e. some features are more salient. Not only are detected features finite and differentially weighted, but our memory for them is even more limited: we can see far more features while they are present than we can remember later.
DeleteAnaïs, please read other replies about learning boundaries along a continuum (rare) vs. learning distinguishing features among many. Recoding refers only to multiple feature-weighting, not boundaries along continua.
DeleteThe continuum of shades of gray, from black to white is like the within-category continuum between two color boundaries. The (relative) abruptness occurs between the two sides of a color boundary. Look at some rainbows, or digitized spectra.
This reading allowed me to understand that categorization was not just way to accommodate for limited memory but that forgetting is fundamental to our ability to abstract from the things that are physically available to us. Without the ability to ignore unimportant features, we would not be able to categorize functionally similar objects or combine features from those things to ground new concepts. Would it be the case that unsupervised categorization would treat small differences the same as larger differences? For example, if a set of pictures varied on two dimensions: being a triangle or square and having size 130 cm squared and 120 cm squared, would an unsupervised sorter be equally likely to sort them on the basis of either dimension? Or can the physical limits of our brains and eyes be considered a type of supervision that make us more likely to categorize based on shape than size?
ReplyDeleteElena, unsupervised learning enhances whatever differences there are. So it can help with category learning only if the big differences happen to be the ones that distinguish members from nonmembers. Without corrective feedback, there’s no indication of what’s a member or a nonmember, hence no indication of which differences are relevant to categorizing correctly.
DeleteThis reading gave me an insight as to what categorization is. It emphasizes the intersection between categorization and discrimination. Categorization is recognizing one single thing whereas discrimination is telling things apart from one another. Discrimination presupposes some form of categorical knowledge as it reposes on how things are different. Categorization is important as it reduces uncertainty. Indeed, any form of information learning ultimately aims at understanding our world better as well as reducing uncertainty of the unknown. Thus, learning is done through trial-and-error over time as well as feedback from the environment to know if the information is accurate.
ReplyDeleteHi Ines, I agree with your comment! Additionally, this seems to highlight an important motivation in understanding our cognitive capacities. As the reading mentioned, behaviourists such as BF Skinner were focused on learning on the basis of behavioural feedback. They were primarily concerned with the “what” of our categorization capacity, in input/output terms. But in the study of cognitive science, our motivation lays within the “how.”
DeleteInes, discrimination does not presuppose categories. You can say whether two things are the same or different regardless of whether they come from different categories, the same category, or no category.
DeleteIt is successfully learning to categorize (by detecting and abstracting the distinguishing features) that reduces uncertainty (about what to do with what).
Sara>/b>, supervised (reinforced) learning is feedback on behavior. The behaviorists did that. What they did not explain (reverse-engineer) was the causal mechanism doing the learning (how).
From this paper, I learned that categorization and cognition are mutually dependent if not strictly identical processes; meaning that categories are not possible without a cognizer- and that cognition is emergent from categorization in a fundamental way. Theoretically, they do appear strictly identical. The trouble seems to be in figuring out exactly how categories are formed and processed and how cognitive abilities rely upon and structure categorization.
ReplyDeleteThe sheer variety of categorization methods and types (i.e. native vs. learned, UG vs. all the rest) points out that categories may not actually do what they do in the same way at all, and that we may be fooled by our categorization filters into believing that they are functionally identical. In conjunction with the Whorfian view of CP, makes the conclusion, that cognition is categorization, a bit confusing. After all, if categorization is cognition, and categories are all that exist, where is the site of change in categorization, and how can categories modify one another?
I suppose I am wondering in what sense cat. and cog. are identical- is it literally meant that there is a strict identity in what we call categorization and cognition, or is there a compositional identity, meaning that categories compose cognition, but there is some sort of difference in structure?
Apologies for the lengthy post!
I was also wondering about the relationship between categorization and cognition. I think in the last sentence, prof mentioned that cognition is all for and about categorization. From what I understand, we need cognition to categorize. It is a fundamental aspect of cognition because one of the central issues in cognitive science is the nature and process of categorization.
DeleteThere are cognitive (i.e., DOing) capacities that are not categorical: continuous actions and procedures like imitation, motor skills (swimming, dancing), similarity judgments. Categorical means discrete (right/wrong action, member/nonmember): approach/avoid -- and, of course naming referents: “X” or “Y” or…
DeleteAs discussed elsewhere in the replies, the same correct categorization can be based on different feature sets. Categories are underdetermined, approximate, and depend on whatever features can successfully resolve the uncertainty about what to do with what kind of thing.
I had the same questions while I was reading the paper. If I understand correctly, cognition is the capacity to do all things we can do, which involves continuous actions (motor skills such as eating) and categorical processes (telling whether something is edible or not). Moreover, categorization is an essential aspect of cognition, and it helps us to perform the right continuous actions (eating the good fruit, not the spoiled one).
DeleteOne insight from this text I found particularly compelling is that all categories are abstract, since categorization necessarily involves selectively taking note of certain features while ignoring others. This makes me wonder what could be the grounds for the distinction between "abstract" and "concrete" categories that we all intuitively seem to have – why, for example, most people would say that "dog" is a concrete category, and "justice" is abstract. I can think of two answers to this question. (1) Perhaps we judge a category as "concrete" if we acquired that concept via our direct (sensorimotor) interaction with the world – or if it in principle can be so acquired –, and as "abstract" if we acquired it via "hearsay" (i.e. through language). (2) Perhaps "concreteness" and "abstractness" relate to how frequently we use that category; categories that are routinely used are classified as more concrete, and those that are not are deemed more abstract.
ReplyDeleteAre there any empirical studies investigating the factors underlying people's concreteness/abstractness judgements about categories?
Gabriel, I think your reply (1) is correct: Abstractness is a matter of degree, with, on the more concrete end, categories that are learned or learnable directly by sensorimotor learning, and, on the more abstract end, categories that are learned or learnable only indirectly, through (grounded) verbal description.
DeleteHow often we speak about or use more concrete and more abstract categories surely depends on what we do in life, physiotherapy or philosophy.
Solovyev, V. (2020). Concreteness/abstractness concept: State of the art. In International Conference on Cognitive Sciences (pp. 275-283). Springer, Cham.
Connell, L., Brand, J., Carney, J., Brysbaert, M., & Lynott, D. (2019). Go big and go grounded: Categorical structure emerges spontaneously from the latent structure of sensorimotor experience. In CogSci (p. 3434).
The most interesting part that I got out of this paper was the distinction between supervised and unsupervised learning. In a supervised learning environment, machine learning models get better at categorizing images “by error-corrective feedback”, through a system of reward and punishment. This system of “feedback-guided trial and error” works by having humans evaluate the correctness of the algorithm’s output to allow the algorithm to accurately learn. If it returns a correct answer, we give positive feedback otherwise, we give it negative feedback. However, I am wondering if these methods of artificial learning could benefit human categorization?
ReplyDeleteIndeed, if we look at the example of chicken-sexing, it is very difficult for us to even verbalize how to tell a male and a female chick apart: “the invariant features are ineffable in that case: They cannot be described in words”. We can’t even tell what makes the important features of a male versus a female chicken, it is an activity solely learned through trial and error. “But in the harder, more underdetermined cases like chicken-sexing, what determines which features are critical?”. It is interesting that Biederman’s experience led him to discover invariants on chicks abdomens which allowed for easier categorization. Despite finding this physical proof, most would say that this level of mastery in such specific activities, is often attained through some sort of human intuition. But I am wondering, if we let an intelligent algorithm run on samples of chicken sexing could it develop an artificial intuition of its own, different from the human man? Furthermore, could we learn from the machines and extract new rules for categorization?
Étienne, I no longer use chicken-sexing as an example because it is done in the service of unspeakably cruel practices. I use mushroom-picking instead. The principle is the same. With supervised/reinforcement learning, corrective feedback from the consequences of miscategorizing (nourishment vs. indigestion) is what guides the detection and abstraction of distinguishing features.
DeleteIn neural nets, supervised learning is guided by feedback from human classification, but in supervised learning done by humans, the environment can provide the feedback from the consequences of correct and incorrect categorization, as in mushroom-picking/tasting.
It is possible that Kayla (T3) can learn to sort edible from inedible mushrooms because she has a supervised neural net doing the feature-learning in her head, guided by feedback from vegetative cues like nourishment or indigestion, but probably there is not yet a learning algorithm that can learn categories at anywhere near T3-scale.
You should ask yourself, though, what “intuition” might mean: If it’s not magic, it should be possible to reverse-engineer it.
Hi Étienne, I am questioning whether intuition plays a role in this. Are these hypercomplex tasks such as mushroom discrimination not just a very abstract form of categorization? As stated in the reading, we might ignore the discrete and almost "intuitive" qualities of a mushroom kind that drive us to categorize it. If we could explain our reasoning by introspection, then cognitive science would lose its driving question: how do we categorize/cognize?
DeleteI do believe that machines could have sensorimotor capabilities that allow them to categorize more accurately than humans. For example, a machine's "retina" could have more densely packed photoreceptors, allowing it to distinguish higher spatial frequencies and perceive finer details that could differentiate the mushrooms. Perhaps we could learn from these tools?
“Vanishing intersections” refers to when there is no feature that all members of a category have in common. If there are two different categories, C and D, and there is no feature that all members of C share to distinguish them from members of D, then how can you tell C and D apart? (How can you categorize them?)
ReplyDeleteThe answer is that if Cs are all either red or green (but never yellow or blue) and Ds are all either yellow or blue (but never red or green), then the feature that Cs all share is not “red” nor “green” but the feature “red OR green.” And Ds are all “yellow OR blue” (or “NOT “red OR green”). No problem for a neural net to learn to categorize them.
So “vanishing intersections” WITHIN C and WITHIN D do not mean that C and D need to be innate. Fodor is mistaken. What is needed is for category-learning is empty intersections BETWEEN C and D: Nothing that is a member of both.
In contrast, in Week 8 and 9, we’ll learn that for learning the rules of Universal Grammar (UG) there really is something missing. (This is called the “Poverty of the Stimulus.”) What’s missing is not an intersection of the features within all members of UG. What’s missing is non-members of UG. The child who is learning languages never hears (or says) anything that is not in UG (hence never gets correction). Children never make UG-errors.
Yet UG errors do exist (an infinity of these non-members, just as there are an infinity of members of UG). Children simply don’t ever make UG errors. So it is concluded that UG is innate. (More later on this.)
I’m really intrigued by the idea of feature selection and weighting about how our sensorimotor systems do not give equal weight nor detect all features. And the brief mention of the analogy between categories acquired through learning and evolution reminds me of an article I read in my other class. It talked about successful perceptions and behavior arise not because the actual properties of the world are recovered from images but because the ones that have been detected and have high frequencies of occurrence accord with the reproductive success of the species and individual. I think that explains why “our visual system abstracts certain features as privileged.” One of the reasons could be for the sake of reproductive success
ReplyDeleteNadila, the key in supervised learning is always success (doing the right thing with the right kind of thing). And what is right (e.g., which mushrooms nourish you and which poison you) is determined by biology (yours and the mushroom’s). So learning to compete during play might be useful for later reproductive success. And this correlation may be inborn, as a result of prior reproductive success for our ancestors. But reproductive success within a single lifetime is too late as the corrective feedback for, say, learning to categorize edible and toxic mushrooms! The feedback has to be more immediate than that!
DeleteWe’ll discuss distal (evolutionary) and proximal (experiential) causal factors in behavior (reproductive and non-reproductive) next week (7) -- but first please do readings 7a and 7b!
ReplyDelete“Under 11. Learning Algorithms” - This might be a bit repetitive as the concept has been brought on from above. However, just wanted to further see if I fully understood what has been said. Based on what we read, categorization has this direct link to learning especially with the specific category of operant learning. In the beginning, categorization becomes described as having an input and the output being different variations of one another.
More specifically, a response on categorization was - “doing the right thing with the right kind of thing, not for the sake of doing it”. This is then where the learning comes into play where one will experience consequences. Hence, is it safe to make a connection with a person trying to recognize which car is theirs in a parking lot (learning to figure out what is his car)? Specifically, we have all been in a situation more than probably once where we are trying to find our car in a parking lot. Often in a busy parking lot, we come across many cars that look like our own - model, color, etc. In this case, the right kind of thing that would matter is choosing our actual car and the wrong thing would be the consequence of choosing the wrong car. However, as mentioned items have features, and so we must use these to our advantage. Where in the case of the car - one could use the license plate number or inner decor to categorize. Overall, throughout life, categorization is beneficial in terms of it’s relation to learning how to identify what will benefit us and what will not.
Maira, remembering where you parked your car is not really category-learning; it’s remembering. Learning which cars are Subarus and which are Toyotas is category-learning.
DeleteTo categorize correctly of course requires remembering the distinctive features, both for Subaru’s and for “where I parked my car today.” -- The latter, though, is just an on-the-fly category, like “when the cat was on the mat yesterday,” useful to do the right thing at the moment, but not a category you would put in a dictionary or encyclopedia.
I found it very interesting when the paper discussed the fact that categorization is never really black or white. Categorizing is indeed not innate but instead learned by supervised learning that involves corrective feedback, and unsupervised learning. It is considered absolute discrimination, when discrimination refers to relative discrimination. An example that I am thinking of when we talk about the blurry line between two categories is the word zebra. It is hard to know in what category we should place it. If we asked a child, they would probably place it under the category of horses, and although a zebra looks like a horse with stripes, is that the right category? “So when we learn to categorize things, we are learning to sort the alternatives that might be confused with one another.” That quote shows how we can categorize things when it gets a bit more difficult: we try to find other objects from which we can discriminate our object of focus.
ReplyDeleteCharlotte, the features that distinguish different categories are approximate, but the categories are all-or-none (or supposed to be!). Something is either a zebra or not a zebra, and there is a right or wrong of the matter.
DeleteThe idea of "categorization as a dynamics system that change in time" is new to me, it makes sense so much sense to identify categorization this way. Adapting and learning then play a role in stringing categorization capacity with perception. I am wondering, as discussed in "16. Abstraction and Hearsay" that recognizing is when we see a kind of thing that we have see before, does it suggest that in order to recognize, we need to have the cognitive capacity to do gist extraction? That is, not only do we need to be able to understand what's encoded, but the cost is to forget the details?
ReplyDeleteMonica, yes, categorization requires learning to selectively detect the distinguishing features and ignoring the irrelevant features. That’s what Funes could not do. The features need to resolve the uncertainty in what to do with what kind of thing.
DeleteUnder learned categories - the idea comes from opening a dictionary. We open a dictionary and we are able to know about what we can call grammatical categories - nouns, adjectives, pronouns, etc. However, the question that stems is “did we know what word goes under which category as an innate ability or did we have to learn it”.
ReplyDeleteI think the answer, in connection to UG, is a mixture of both. Specifically, growing up we learn the language around us and more or less are able to pick it up quickly (starting around the age of 1). Now we do in fact learn the language, however, based on UG we “have the language acquisition device, i.e., the innate ability to be prepared to learn the language. In such cases, we have the innate ability to begin learning language from predisposed grammar rules within us which are the “identical features” shared among languages.” (took my reply earlier). Here, the predisposed grammar rules involve nouns, adjectives, etc. Thus, based on UG, I think it is a matter of knowing about these categories innately and eventually applying them when we begin learning a language.
Apologies, just saw your reply to what UG is above! So this answer may be invalid!
DeleteMaira, the UG/OG distinction will come more into focus in Weeks 7 and 8. Meanwhile I hope the replies here will help.
DeleteThe distinction between unsupervised and supervised learning was slightly perplexing to me. Unsupervised learning does not have an error corrective feedback mechanism and is likely used to distinguish categories that are very different from one another, where there is only one way to separate the categories. On the other hand, supervised learning does provide us with a corrective feedback mechanism that tells the system whether it is doing the right thing with the right kind of thing... given that it does it in the right context. So, in order to understand the context, it appears that a supervised learning mechanism must have some kind of innate knowledge in order to guide a system while it is learning. If we are supervised learners, this means that our brains must have some innate mechanism capable of understanding the context in which we are categorising and providing us with feedback to correct any errors that we are prone to making as we categorise. Is it likely that we perform this task unconsciously? As we are not always fully aware that we are performing this at a given moment.
ReplyDeleteSara, unsupervised learning is sensitive to feature/feature correlations in the input data; supervised learning is sensitive to correlations between input features and right or wrong outputs.
DeleteThe context of inter-confusable alternatives that category learning attempts to resolve is not innate; it is sampled when an organism is learning. It can be representative or non-representative. It can grow; and distinguishing features can be revised. Both categorization and features are context-dependent and approximate.
Different contexts require different features to distinguish what to do what with.
Now, can you explain that to kid-sib?
The first reading of the week provided a very holistic explanation of categorization while linking it to cognition. As stated in a few other skywritings above, I have previously learned about categorical perception in previous psychology classes, but without having considered the innate nature of some categories, and how others are learned. The importance of error correction within categorical learning was quite interesting as well, as it prevails in context-dependent categories which may lack salience. This article also mentioned abstraction as the ability to focus on a subset of a sensory input, which is an important skill for categorization, and was stated to be categorization in the reading. Prior to reading this article, I had a much more surface-level understanding of categorization, without necessarily seeing its relation to cognition. However, it is now clear to me that the categories we are able to make (which determine our view on the world) are influenced by our language, truth, affordances, etc.
ReplyDeleteThe reading focuses on how we learn and categorize through our sensorimotor systems. What our sensory motor systems "afford" entails what information they can gather from the external environment. But what we gather from the external environment is also shaped by our systems such as in the boomerang example whereby we change our position in view of the boomerang, but our systems extract the invariance to present visual constancy of the object. The article demonstrates that categorization is both an innate and learned process. Learning can be either unsupervised, such that categories can be identified without any feedback, and supervised whereby categorization is correct or confirmed by an error feedback.
ReplyDeleteAn example of unsupervised learning is after being repeatedly exposed to horses so I am able to categorize all horses I will perceive in the future as horses through only internal analysis. Whereas, supervised learning would mean that upon exposure to a monkey, I categorize it by a horse, and my friend (or any provider of information) corrects me that it is not in fact a horse. From now on, I am able to recognize that a monkey does not follow under the category of horse.
Laura the difficult cases of category learning require a lot more trials and error-correction before your brain detects and abstracts the distinguishing features that ensure you do the right thing with the right kind of thing.
DeleteFor me, this paper clarified why categorization is such an important aspect of cognition - if not the entirety of it! Categorization occurs when we do the same thing with the same kind of input. Most categories are learned through exposure, trial-and-error with feedback, or via verbal instruction ("hearsay"), although some categories, like UG, are innate.
ReplyDeleteAll categories have an extension - the set of all members - and an intension - the (set of) necessary feature(s) of a member - both of which vary with the demands of the context. For categorization to be possible, we must have a way to detect and use the features of an input that afford correct categorization. Features are weighted differently depending on the demands of the context, thus, we selectively forget "unimportant" features.
Although it seems like we form "concrete" categories from sensorimotor experiences and "abstract" categories through "hearsay", since sensorimotor experiences are all abstractions and all categories can be traced back to their sensorimotor origins, even "concrete" categories are "abstract".
If categorization is cognition, will reverse-engineering something that can categorize like a human solve the "easy problem", namely "explain how/why we can do the things we can do"? In your opinion, professor, are artificial neural nets our best bet for achieving this?
Polly, good summary.
DeleteAll categories are abstract. (How?)
Cognition is categorization, but not just categorization. See continuous sensorimotor skills, imitation, and relative judgments in the comments and replies.
Neural nets are the first serious model of feature-abstraction in category learning -- but not the last.
If I'm understanding the paper correctly, I believe that all categories are abstract because we categorize a thing by selectively abstracting its invariant features. The "invariance" of a feature is determined by what the confusable alternatives are in the given context (i.e., feature invariance is relative to the variance of confusable alternatives). According to this line of thinking, all categories are the result of selectively abstracting some features and ignoring others. Therefore, all categories are abstract.
DeleteCould the way in which Fodor is wrong about innate categories using the vanishing intersections example be thought about in terms of weights of instances in a feature space? If I were to experience multiple different instances and create vectors between instances based on what co-occurs together, I would be creating categories based off of co-occurrences as well as shaping the feature space to push things that don't occur together in a specific context closer together as well.
ReplyDeleteOn this note, I was wondering about how exemplar theory (categorization theory whereby a category is represented by all of its members and an instance is categorized based on its total similarity to all category members) relates to neural networks and category learning. Do the theories behind neural networks overlap with exemplar theory since instances are being related to one another in a feature space based upon co-occurrneces? Is this just the unsupervised learning part of neural nets? If that is the case, I am not sure if I understand how the supervised learning occurs in neural nets, if unsupervised learning is based on correlations between instances. Does supervised learning occur through whether or not output based on categories made through co-occurrences matches the correct output, and then changes the weights between members based on its correctness?
ReplyDeleteCategorization works as a filter to reduce information content in the environment by marking certain things as analogous and grouping them together. This process is continually updated based on our learning through new sensorimotor experiences. We can acquire categories directly, through unsupervised learning (passive exposure without feedback) or supervised learning (trial-and-error), or indirectly through hearsay (verbal transfer of a category based on other already-grounded categories).
ReplyDeleteOur perception of the environment and how we interact with it is shaped by the way we categorize things. Categorization allows us to give meaning to the world. This is exemplified by Borge's Funes, who remembered everything and therefore did not have categorization capacity, so “different instants were incomparable, incommensurable.” Thus, without categorization, nothing would have meaning.
I do agree that category gives us meaning to an extent. Saying categorization works as a filter reminds me of Broadbent’s early filter model and Treisman’s Attenuation Theory. Where both theories focus on filtering out information to the point only the information that is important/relevant will be processed for meaning. I guess what you are saying about categorization is similar.
DeleteMoreover, I think, although not sure if valid to say, is the role of visual areas which process categories of information to categorization. For example, the extra-striate body area has the role of process and then categorizing non-facial body parts; the parahippocampal place area has the role to process and categorizing conscious recognition of places and scenes we encounter. If categorization works as this filter then I wonder how exactly it is done through these parts. But then again body parts having these roles in aspects of categorization I guess leads to the question of how can they categorize body parts from body parts and place to place - which brings us back to the question of how we can do.
Hi Maira, I agree that it is interesting to consider how different brain areas contribute to our perception, and thus to the categories we use. Many neuroscience observations can contribute to the tenor supported by Prof Harnad that categorization is primarily learned: research has shown that face recognition areas in our extrastriate visual cortex adapt to certain faces and fire increasingly when shown the face again. All humans have this underlying mechanism - our brains are selective to certain features, which helps us understand how we abstract key features similarly.
DeleteHi Darcy,
ReplyDeleteUnsupervised neural networks find clusters and correlations among unlabeled data. Supervised neural networks are trained using labelled data through trials and errors, allowing this network to give desired outputs with an input. Both are machine learning models that can be used for various goals. The exemplar theory is a theory on how humans categorize, and this theory can be reconstructed by neural network models, such as in this paper: Comparisons of prototype- and exemplar-based neural network models of categorization using the GECLE framework https://escholarship.org/uc/item/5s86d3cp.
Hope this answers your questions!
I found the explanation of language as a means to indirectly learn categories via hearsay to be helpful, especially in terms of concepts like truth and beauty, that were discussed. If categorization is defined in terms of abstraction from sensorimotor experience and supervised correction, it feels like a natural progression to ask what that means for more “abstract” categories (though all categories are abstract, as mentioned above). Hearsay is helpful in that regard, I think especially for categories like feelings. Love, for example, is popular in culture for being particularly difficult to categorize (particularly in coming of age movies). It leads to quite a lot of hearsay that attempts to ground explanations in sensorimotor experiences—like feeling out of breath.
ReplyDelete
ReplyDelete“What the stories of Funes and S show is that living in the world requires the capacity to detect recurrences, and that in turn requires the capacity to forget or at least ignore what makes every instant infinitely unique, and hence incapable of exactly recurring.”
I'm not sure I understand this quote completely. According to my understanding, it is essential to note what is frequently happening in our lives. And in order to do so, we must first determine what makes each event "unique." However, in this case, does this quote potentially contradict theories such as SHY and sleep to forget, sleep to remember? -> where we want to be able to remember relevant stuff (which can be analogous to “unique”)?
I apologize and take this back. I just remembered we are focussing on aspects that are cognitive. Therefore, since these theories focus on vegetative states rather than cognitive it makes them irrelevant to this quote.
Delete6a: Categorization, same as consciousness, cannot be explained through introspection. We can share the category we generate for a given input, but we are once again blind to our inner discrimination mechanism - the “how”. The process depends on the absolute judgement of an object; the features on which we rely to sort the object are abstract - they will depend on the nature of the objects it is being contrasted against, which will determine what features matter in the context. In my opinion the point 27 on the limits of hearsay was important in this text to tie categorization to cognition by exemplifying that we cannot explain how we categorize.
ReplyDeleteI never truly realized (or appreciated) how much categorization plays a role in shaping the world we perceive. Furthermore, Whorf's hypothesis intensified my interest in categorization even more. However, every human and, to an extent, every living being has similar cognition levels. Our cognition methods to categorize may indeed rely upon the language we have been taught (or the one we have naturalized within ourselves). But there is a problem that I can see with this approach. For the Whorf hypothesis to take effect, one needs to only think of things in words. This is fine, but as the paper has stated, most of our categorizations come from the abstractions we make about things. And we cannot abstract stuff with words.
ReplyDeleteThe abstraction and amnesia section was really intriguing about the short story, “Funes the Memorious,”. Specifically on how having an infinite rote memory would only be debilitating rather than a useful tool because it prevents an individual from being able to consolidate their long and short-term memories. Hence it would prevent them from being able perceive their environment and be an active member because they are not able to forget or engage in selective forgetting. The information on recurrences not being able to processed because it would require being able to understand abstractions but Funes would not have been able to do that because his memory was essentially full.
ReplyDeletePrevious psychology courses I have taken have spoken about categorization in the context of social schema development. However, neglected, apart from a general discussion of assimilation and accommodation of information, has been the mechanisms of category formation. This article, through explaining role of passive exposure/unsupervised learning, active/supervised learning or heresy, developed my understanding of how this is done.
ReplyDeleteConnecting this to the concept of developing a conscious AI, just we take in sensorimotor information, the computer would also require the ability to take in information from the world around it.
As polly previously suggested, neural nets could be a way for computer categorization to occur using this information. I do have a few questions regarding this however?
Would the neural nets have an underlying processing framework from which they would begin integrating information together?
and do we as humans have an agreed upon underlying framework that would be analogous?
What stood out to me was the idea of visual constancy wherein our visual system is able to perceive a shape as invariant regardless of its motion which is due to our assumption of what the object should look like or how it should behave. This is interesting because it made me think of optical illusions where our visual system is tricked into perceiving something as moving, or that one part of a picture is a different shade than another part even though they are the same colour. In many ways, our visual system tries to optimize on informational efficiency through assumptions formed through past experience wherein we place certain occurrences into categories that we later attempt to apply. However, sometimes, these optimizations can distort our perception of the world.I feel like this is one of the areas where the goals and outcomes of our cognition is clearly expressed (also shows how the visual system is adaptive and dynamic).
ReplyDeleteLearned CP effects are demonstrated through the experiment involving learned texture categorization. The experiment found that there was no difference before and after learning for the easy categorization task, but the hard categorization task resulted in within-category compression and between-category separation after learning. This means that after learning the hard texture categories, between-category differences looked bigger than within-category differences, while the same was not true for the easy texture categories. So, only the task in which subjects had more difficulty learning to categorize the textures resulted in CP.
ReplyDeleteWhy is this the case? Is it because when we learn harder categories, we are learning to sort out the alternatives that might be confused with one another, whereas there is little confusion with alternatives for easier tasks? In that case, wouldn’t the presence of additional examples of easy-category textures still help us cement the category through within-category compression?
Although it was not the main focus of the reading, the section on Chomsky’s Universal Grammar and its relation to innate and learned categories raising an interesting notion about the connection between categorization and language. Despite the flaws in Fodor’s “Big Bang” theory of categorization capacity, it is clear that we do have some sort of innate ability to form categories of things in our mind, and this ability of categorization exists independent of language, or rather the language that we learn when we are children. For example, even though the concept of nouns, verbs, adjectives etc. is not the same in Chinese as it is in English, children that speak Chinese will still grow up to learn the categories of “bird” and “mammal” and the differences between them, implying that categories really do form primarily via sensorimotor interactions with the world.
ReplyDeleteThe story of Funes, who due to having an infinite rote memory and relatedly the inability to forget lives every experience as new and unique, reminded me of the famous real-world case of Patient H.M. Having undergone hippocampal surgery to treat his seizures, H.M. lost the ability to form new conscious memories relating to new lived experiences (declarative memory in more technical terms). For people who H.M. only met after his procedure, each meeting was as if he is meeting them for the first time - like Funes. However, for Funes the issue is the capacity for "perfect" memory, or the inability to forget; and for H.M., the issue was the inability to form new memories. It is interesting that these two changes should converge to create the same sorts of behavioural differences in these cases.
ReplyDeleteI found the comparison of Chomsky’s universal grammar with Fodor’s innate categories to be helpful in illustrating how far Fodor goes in his claims, and how wrong he is. The key difference being there is no evidence to support an all around “poverty of stimulus” argument for category learning. So since humans have to learn categories it makes sense to reverse engineer that capacity. I think category learning is a particularly good example of a function which can be better explained by reverse engineering capacity in machines than by studying different areas of the brain. The conclusion that cognition is categorization is bold but one that I think makes more sense than past attempts of defining cognition (ie cognition = computation).
ReplyDeleteMy skywriting was removed so I am posting it again:
ReplyDeleteI am not sure whether I understood the “Vanishing Intersections” section of this reading. When talking about words such as “truth” or “beauty”, do we have an innate mechanism to categorize those words because they are abstract words, i.e., we cannot detect them with our sensory motor system and therefore there is no room for trial, error, and feedback? And what is exactly meant by the “intersection” of content words?