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
NOTE TO EVERYONE: Before posting, please always read the other commentaries in the thread (and especially my replies) so you don't just repeat the same thing.
ReplyDelete**BLOGGER BUG**: ONCE THE NUMBER OF COMMENTS REACHES 200 OR MORE {see the count, at the beginning of the commentaries] YOU CAN STILL MAKE COMMENTS, BUT TO SEE YOUR COMMENT AFTER YOU HAVE PUBLISHED IT YOU NEED TO SCROLL DOWN TO ALMOST THE BOTTOM OF THE PAGE and click: “Load more…”
Delete________________
Load more…
________________
——
After 200 has been exceeded EVERYONE has to scroll down and click “Load more” each time they want to see all the posts (not just the first 200), and they also have to do that whenever they want to add another comment or reply after 200 has been exceeded.
If you post your comment really late, I won’t see it, and you have to email me the link so I can find it. Copy/Paste it from the top of your published comment, as it appears right after your name, just as you do when you email me your full set of copy-pasted commentaries before the mid-term and before the final.
——
WEEK 5: Week 5 is an important week and topic. There is only one topic thread, but please read at least two of the readings, and do at least two skies. I hope Week 5 will be the only week in which we have the 200+ overflow problem, because there are twice the usual number of commentaries: 88 skies + 88 skies + my 176 replies = 352!. In every other week it’s 2 separate topic threads, each with 88 skies plus my 88 replies (plus room for a few follow-ups when I ask questions.
This isn’t my official skywriting for the week; I was just wondering if someone could further explain Figure 1. I don’t understand what is meant by “hidden-unit representations”, or what determines their placement in the 3-dimensional space in the bottom two cubes. Thank you in advance!
ReplyDeleteThis picture is incomplete. The same object or input could belong in many different categories, depending on what you have to do with it in different contexts. So learned CP must be weaker, more context-dependent and attentional than innate CP. But the effect is similar. And neural nets don't just learn categories, by detecting their features, they also produce the CP (separation/compression) effect.
DeleteHow? Nets are not sentient. They cannot perceive. But what the linked image at the top shows is that within the net, the internal representations" -- not weasel-words in this case! of the members (blue) and and the non-members (purple) become more and more separated between the categories, and compressed within the categories, as the learning progresses. This is all a result of the changes in the connection weights in the network during the course of supervised learning.
Think of each blue input as the position, in N-dimensional hidden-unit space, of the members of the category and each purple input as the position of the non-members. At the top, before the supervised learning, the blues and purples are all overlapping and inter-confusable. At the bottom, after the learning has been successful, they are maximally separated.
Bonnasse-Gahot, L., & Nadal, J. P. (2022). Categorical Perception: A Groundwork for Deep Learning. , 34(2), 437-475.
DeleteCould the “Ugly Duckling Theorem” be seen as an example of learned CP? In this case, the words of the category of ugly ducklings’ explicit description are first and foremost grounded in sensorimotor invariants. Then, given a group, we attempt to categorise each as a member or non-member of the category of ugly ducklings. After sufficient trial-and-error feedback, the relevant features are privileged (i.e., up-weighted) over other features that also differ between members and non-members that are not critical to the categorisation of ugly ducklings (i.e., those that do not particularly differ along member/non-member lines are down-weighted).
DeleteThe fact that the phenomenon of detected features being finite not only impacts our memory as well but to an even bigger degree, I wonder if this provides further insight on the categorisation problem, namely what we do detect (and remember detecting).
DeleteJocelyn, thoughtful comment.
DeleteTo make it more realistic, there would need to be something to DO that depended on distinguishing the cygnet from the ducklings. For that, the innately upweighted salience of size and color would already pre-solve the problem (as Watanabe himself says: it's an evolved, built-in bias already).
A bigger challenge would be "always pick the southernmost one".
Remember not to conflate category-learning and CP. Category-learning here would be the trial and error detection of the distinguishing features of the "southernmost bird" category.
CP would occur if that learning produced a feature-detector that made the southernmost bird pop out (because of an up-weighted feature, producing a between-category separation effect [and perhaps also a within-category compression effect] in the perceived similarity among the birds). CP does not happen with every learned category. Only the hard ones, that need to pop out clearly and quickly.
This is just an artificial example, but it's essentially the same as the mushroom island: If the distinguishing features are detected, abstracted and up-weighted, the edible mushrooms should pop out.
This comment has been removed by the author.
ReplyDeleteDo them in the week after tomorrow's lecture -- as early as possible, so I have time to reply.
DeleteThis article makes the interesting distinction between direct sensorimotor learning through trial-and-error experience (implicit knowledge) and indirect learning through 'hearsay' (explicit knowledge). Language seems to have a powerful adaptive advantage by allowing us to transmit knowledge and skills faster, but I'm pretty sure explicit learning is not always the best approach. What implications are there for education? In traditional education, it is typical to ensure students have as much of an explicit understanding as possible.
ReplyDeleteWhile our first language is implicitly learned at an early stage, the second language is usually taught later at school through an explicit approach: by learning grammatical rules and memorizing vocabulary through translation into our first language. I remember that the best students were those who had a good intuitive knowledge (by watching movies in the second language) but a pretty poor knowledge of grammatical rules. This explicit approach to language turned out to be personally counterproductive since intuitive and explicit knowledge were sometimes contradicting and competing. To my mind, it might be better to gain an intuitive understanding at first, and then learning the explicit rules will be easier instead of asking people to work out the structure consciously from the beginning. The same could be said for sports: I think that people should be able to develop their skills through practice before explicitly teaching them the optimal techniques, which are sometimes harder to perform.
From an educational perspective, starting with implicit learning through trial-and-error experience would involve students more actively and motivate them. Afterwards, acquiring the explicit knowledge will be easier and will give them a fresh perspective. This, I think, is the approach supported by a flipped classroom.
The two means of learning new categories are (1) direct sensorimotor learning from experience and (2) indirect verbal learning.
DeleteThat doesn't quite align with implicit vs. explicit learning. Verbal learning is always explicit, but sensorimotor can be both:
You can learn the distinguishing features of a category directly, without being able to identify or verbalize the features (as with Ohri's 3rd grade school-teacher); or you can learn them directly, but also identify the features verbally.
To learn categories indirectly, the words referring to their features have to already have been learned. They have to already be in your vocabulary.
A lot of formal education is verbal, but it has to be grounded in prior direct experience; and direct experience remains a useful complement, where possible, all along the way. Kid-Sib always prefers -- and prefers to learn from -- examples rather than just words (showing, rather than just telling).
Language learning, especially 2nd-language learning, as well as motor skills, are not just category learning. Observation and imitation (and mirror capacities) are important too.
Explicit verbal instruction is not the natural or the most effective way to learn a first language; it helps more with a second language, but there too, as a supplement rather than a substitute. And you can't learn to swim or play tennis from just a textbook.
Imitation may be mostly implicit, but bear in mind that trial-and-error need not be: It can be partly explicit hypothesis-testing ("Hey, maybe the striped ones are edible"), all going on in your own head, without an instructor, but deliberately, and (as in long division) with awareness that you are doing it.
In to cognize is to Categorize: Cognition is Categorization, one notion that was emphasized was the one of abstraction and its role in categorization. Categorizing being an absolute judgment to determine if an object is a member of a category requires both to be able to identify similar elements and to be able to discriminate between that object and another member of that category. Because a category is a kind of things, us humans need to find the balance between recognizing two objects as completely different but still similar enough that they would be of the same kind. So, through abstraction that singles out a subset of sensory inputs and ignore the rest you are able to select the features that are important for categorizing two objects under the same category and ignore the features that would discriminate them too much to the point where they wouldn’t be recognized as being of the same kind. I personally find it fascinating that us adults do it constantly without realizing how meticulous the task actually is, so how confusing it must be for a child to learn that two computers are computers but not the same one if they look the same/similar.
ReplyDeleteI also enjoyed the portion of “To Cognize is to Categorize: Cognition is Categorization” that spoke about abstraction. The part that relayed the fictional story of Funes(who had a perfect rote memory), and the true story of ‘S’, who was not fully able to remember everything he ever saw, but had a significantly better rote memory than most, was particularly helpful in developing my understanding of the importance abstracting the sensory inputs. I never thought that having an extreme memorizing ability would be a handicap, but this section nicely highlights the importance of abstraction, as viewing every moment and object in time as innately unique, limits our ability to make categories.
DeleteUnrelated Side-note: In class, it was asked to take a page of a dictionary and observe how many words referred to individuals vs. those that are categories. I found that of the words on the page 14/37 could refer only to individuals(which is around 38%, indeed lower than 50% as suggested in class).
Learning to categorize is learning the right thing to do with the right kind (category) of thing: "Is this the kind of thing I do THIS with or THAT with? To be able to do this you need to find the features that distinguish members of the "THIS" category from the members of the "THAT" category. So your brain detects and abstracts the distinguishing features and ignores the irrelevant features.
DeletePlease always read the replies to other comments so they nee not be repeated.
I also found the discussion of Funes and S and its lessons on abstraction and forgetting to be fascinating and insightful. In particular, the importance of selectively forgetting and ignoring in the process of abstraction stood out to me, as it allows only the relevant features to be used to categorize. This was further illustrated by the ugly duckling example, where to categorize the swan chick compared to the ducklings, the features of positionality would have to be ignored, and only the features of color should be considered to correctly categorize. Otherwise, if all features are considered, as is the case for Funes, everything would be unique and impossible to categorize.
DeleteThe use of abstraction and selective ignoring in categorization also made me think of ideas such as heuristics and system 1 thinking because of the rapid, instinctive nature of selectively choosing certain features to focus on.
Omar, I think Borges was an ingenious "cognitive scientist", adding a profundity to otherwise often banal-sounding ideas.
DeleteSomething that I found really interesting about this article is the manner with which it addresses the problem of sensorimotor information’ integration. We know that each sensory organ is triggered by an environmental stimulus which constitutes the somatic system, but one problem remains: how are we able to perceive what is in front of us by integrating pieces of information encoded in electric stimuli that travel through the form of action potentials to our central nervous system? This article describes mechanisms such as the categorical perception or reinforcement learning that show our innate abilities to categorize objects in the environment and to act consequently to it (even though we are not always aware of why we do it that way as shown in The adaptive advantage of language: Hearsay) , but that’s just based on what we think is the correct action pattern in response to the stimulus, such as the pecking behavior of the bird to which we present 2 targets and a color. What is, for me, really interesting about this article is that it goes beyond the behaviorists’ point of view and addresses the “black box” issue.
ReplyDeleteWhat addresses the "black box" issue is reverse-engineering, computer modelling and Turing Testing.
DeleteThis reading summarizes how we use are sensorimotor capacities to categorize, however there was a part that I was stuck on. In part 21: Recording and feature selection, I didn’t understand what we meant by increasing our capacities to categorize, does this mean more specific categories?
ReplyDeleteAlso, from other courses and from this reading, I took that categorization is something innate, or automatically done. However, throughout the reading, trial-and-error learning is mentioned a lot, and I do understand that we learn through errors and feedback, however if we can —from feautures and sensorimotor inputs— distinguish things innately, where does reinforcement learning come in the picture?
In this context, I think "increasing our capacities" means improving or even expanding our ability to make categorizations.
DeleteAs for your second question, the idea of categorization being innate is related to our basic ability to classify things. However, we rely on trial and error learning to improve and expand our categories. By learning from the consequences of our actions, reinforcement learning allows make our categorization more adaptive.
My understanding of categorization being innate is that, we are born with the categorization ability, as in we can categorize kinds of things. But it is from learning (either trial-error or hearsay), that enables us to DO actions with the right kind of things.
DeleteCategorization: To categorize (things) is to do the right thing with the right kind (category) of thing. Categories have members (things that are in them) and non-members (things that are not in them). Apples are members of the category "apple" and billiard balls or tomatoes are not. We eat apples, but not billiard balls.
DeleteFeatures: Categories are distinguished by their features: Apples are soft, billiard balls are not. For some categories, like colors, we are born with innate feature-detectors. We only need to learn their names in our language. For categories that we need to distinguish, such as apples or billiard balls or red blood cells, we need to learn their features. (What are the three ways to learn them?)
To answer your question Dr. Harnad, the three ways to learn the features to distinguish categories is through unsupervised learning, supervised learning, and verbal instruction (hearsay).
DeleteUnsupervised learning, due to its passive nature, would only be able to correlate the frequency that certain features appear in conjunction with other features and not actually distinguish any categories.
Supervised learning, on the other hand, could distinguish categories because through error-corrective feedback one could learn to do the right thing with the right kind of thing, and thus correlate features with specific outcomes.
Verbal instruction is another way to learn features since the distinguishing features of the categories, and the categories themselves, can be conveyed through language (assuming all the words used are categories grounded by the speaker and hearer).
I was going to write an amendment of Ishan's statement above that unsupervised learning cannot distinguish any categories but as I thought about it I'm not sure that I'd be right to correct it. I wanted to point out that unsupervised learning is only inadequate when the categorization is context dependent. If there's only ever one way to cluster the same sensory shadows and it doesn't matter on the context then unsupervised learning would suffice. But then I wondered whether anything is ever context independent and I'm not sure that there exists such a thing. In this case, is all learning supervised? The paper makes it seem as though unsupervised learning does a bunch of the grunt categorization (feature grouping?) and whatever cannot be accomplished by this is left for supervised learning through trial and error. Is this the right way to think of it?
DeleteCategorization helps solve the grounding problem by establishing a connection between symbols and sensorimotor experiences. In the reading, Professor Harnad introduces algorithms as potential mechanisms for understanding the process of categorization. However, if these mechanisms are algorithmic and, therefore, computational, how can they solve the symbol grounding problem? While I understand the concept of categorization, I'm still uncertain about how this process bridges the gap between symbols and their real-world referents.
ReplyDeleteThe solution to the SGP is hybrid, computational and analog (including sensorimotor function). Neural nets (which can be implemented either computationally (as a digital algorithm) or analog (as real parallel and distributed neural networks) are the components that can learn to detect and abstract the sensorimotor features that distinguish members from nonmembers of a category.
DeleteWhen we are learning a language, what are we really learning? I am tempted to answer: we are expanding what we can do with the recognized categories. But the question goes beyond just language. As the article finally states, cognition is categorisation. Therefore, everything we learn is to expand our capabilities when interacting with categories. In which, as we discussed in class, animals also perform categorisation, meaning they are all cognizers. What animals cannot do, is propositional thoughts. Is this because they do not have language? (Or is the causal relationship the other way around?) Speaking of language development, could it be that humans have UG and animals don't?
ReplyDeleteTina, cognition is mostly categorization. But continuous motor skills (like swimming or dancing) are cognitive too, though not categorical (except at certain decision points).
DeleteYes, language is a (revolutionary) extension to our ways of learning as well as interacting with category members. Nonhuman animals can of course also learn categories, but in only two of the three ways (see below), not language (verbal instruction).
There is no 4th way (as far as I know). Imitation is good for learning continuous skills (mirror capacity) but not for learning categories' features, except if you want to consider imitation a dynamic form of unsupervised learning--which is likewise not enough for most category learning. (Why not?)
Hi! I am also curious about this question when I read the article. Section 13 and 20 has revealed that the key questions of categorization in machine learning are how to sort things correctly and what is the winning feature. Supervised learning helps more on finding the winning feature through feedback-guided trial and error training. Especially it seems like unsupervised learning is highly input-dependent so that it would have problem on sorting things correctly.
DeleteEvelyn: And also, the same things can be sorted in many different ways, depending on what you may have to DO with them. So the distinguishing features can be very different ones. Unsupervised learning can only highlight feature/feature correlations. Supervised learning highlights feature/outcome correlations,
DeleteYour points as well as your guesses are all correct.
ReplyDeleteFrom what I understood from the reading, there are 3 ways in which we can learn and detect categories. There is unsupervised learning in which we learn through exposure to a category, and we learn to detect correct features. Then there is supervised learning which involves trial and error with feedback. Lastly there is verbal learning, which involves verbal hearsay and therefore requires the use of language which is an exclusive human capacity. After finishing the reading this got me asking myself the question of how come that even if only humans have language, other species seem to know categories. Indeed, in the video associated with this section professor Harnad mentions that even pigeons can learn categories. But this makes sense because knowing categories is an adaptive feature that determines how we see and act upon the world. For example, Zebras know how to recognize what a lion is, they see it as another animal, but they also know to associate it with danger. They can categorize because this is something that is crucial to their survival, and they form these categories through supervised and unsupervised learning.
ReplyDeleteValentina, mostly right, but here are a few corrections:
DeleteYour model for supervised learning should be the mushroom island.
Unsupervised learning means passive exposure with no feedback as to what is a member or nonmember, or even that there are categories. (Hard question: with unsupervised exposure, how could you know how many categories there were, let alone which was which?)
How does unsupervised learning find features? (Hint: it is based on correlations and frequency; but what with what?).
[What does the "verbal hearsay" say? Think what that might be, before reading further below.]
Language (i.e., propositionality) allows a speaker (or dictionary or ChatGPT) to tell (i.e., describe to) a hearer the features of a new category, C, that the speaker already knows but the speaker doesn't yet know.
To do that, both the speaker and the hearer need to already know the referents of the category-names of the features of C, but only the speaker needs to already know that those are the features of C.
Don't confuse "supervised learning" with learning by "instruction": (What is the difference?)
("Supervised learning" is just a "batched" form of reinforced learning, which can be done all at once, instead of serially, trial by trial. Its defining feature is positive and negative feedback indicating success or failure in a trial. On the mushroom island, the only supervisor is the metabolic consequences of eating the right or wrong kind of mushroom.)
This article described multiple ways of learning categories. Supervised learning relies on feedback. Unsupervised learning relies on structural similarities and correlations. Hearsay is learning through language, but it requires that fundamental concepts are grounded through the senses for the learner.
DeleteAttempting to answer some questions above. With unsupervised learning, one could know how many categories there are through observation and extrapolation of possible features. With enough exposure, one could start to make connections of what seems similar and what does not. And which features similar things share. Every new item can be matched with previously determined features. For example, with mushrooms one could think of colour, texture, gills, and shape. One would however not think of spore prints or dna sequencing through unsupervised learning. One may also fail to include mushrooms that are not the standard shape (stem and cap). One may also fail to notice that mushrooms are a category at all, they may seem like plants. It seems to me that through passive exposure it would be very difficult to learn all possible categories.
Nicole, since you have to DO things to be able to know whether you've done the right thing, just passive exposure, without active trial and error and feedback from the consequences, can only help in categorization if the features that are detected happen to turn out to be correlated with doing the right thing, once you start trying to do it! The features that stand out from passive observation could just as well turn out to have nothing to do with the features needed to distinguish the edibles from the inedibles.
DeleteI found the emphasis placed on the functionality of categorization in Professor Harnad’s paper particularly interesting. Throughout this paper Harnad highlights the importance of categorization for what humans are able to do, both by demonstrating what cannot be done if the ability to categorize is missing (as demonstrated by the examples of Funes and ‘S’) and by emphasizing the myriad ways we accomplish categorization and the variety of tasks it enables us to complete (such as sustaining ourselves on a mushroom island, sorting newborn chicks, and communicating this information to others). This paper, and discussions from the previous two lectures, have foregrounded this process as necessary to help us solve the easy problem (how it is we are able to do what we are able to do). Further, it has led me to wonder how this dynamic process could be implemented or reverse-engineered in a T2/T3 machine, and what implications this would have for the machine’s ability to pass its respective TT.
ReplyDeleteShona [I'm saying everyone's name now to help me stamp the face, the skies and the identity of each of you in my brain when I have to integrate a composite mark for the mid-term]:
DeleteYour summary's fine. What's not categorization? (See other replies).
All organisms are machines: That just means they are causal mechanisms.
And I've stopped using the chicken-sexing example: why?
When defining learning, the presupposition of there being such a thing as an error briefly came up. I found this side mention very interesting because this, to my understanding, needs to be made explicit due to the context-dependent nature of categorisation (i.e., the very same things can often be categorised in many different ways). This also comes up when categorising more “abstract” objects that are not directly accessible through our senses (e.g., whether a verdict is just) or more based on personal sensory tastes (e.g., whether a dish is tasty).
ReplyDeleteUnrelatedly, I was wondering if someone could explain this first part of part 29 to me as I’m not really sure what it is supposed to convey: “Mere cognitive scientists (sensorimotor roboticists, really) should not presume to do ontology at all, or should at least restrict their ontic claims to their own variables and terms of art -- in this case, sensorimotor systems and their inputs and outputs. By this token, whatever it is that “subtends” absolute discriminations -- whatever distal objects, events or states are the sources of the proximal projections on our sensory surfaces that afford us our capacity to see, recognize, manipulate, name and describe them— are all on an ontological par; and subtler discriminations are unaffordable.”
Hi Jocelyn! Hope I could provide any help.
DeleteDuring the latest lecture, I remember professor remind us metaphysics will not be discussed in this course. It is hard to figure out what is the truth and the reality. For instance, we could not explain other-mind problem perfectly unless we could read others' mind. Human beings has limited sensation: many birds can see into the UV, which is outside the visible spectrum of humans. For this reason, we only sense part of the world, and if we are stick to what is the essence of things and other ontological question, it is absolutely a dead-lock there. (Welcome to correct if I misunderstand anything!!!)
Jocelyn, good summary. Evelyn, good reply to Jocelyn (about that horribly, unkid-sibly passage I am ashamed to admit I wrote 7 years ago!). Yes, we're trying to reverse-engineer what people can discriminate and categorize (i.e., what they can do with their sensory and motor capacities), not about "what there is" in the world. That's the business of physics and metaphysics.
DeleteIn the reading by Harnad “To Cognize is to Categorize: Cognition is Categorization”, categorization is defined as a sensorimotor system which must be learned. However, categorical perception, or the ability to detect a sudden perceptual change at a boundary of two dimensions is described as being innate. Categorization depends on selectively attending to some features, while ignoring others, in which our sensorimotor systems not only assign different amounts of importance to different features, but they do not even detect all features. The detected features, some more salient than others, are finite, and our memory of them are even more finite. In other words, we detect many features and remember fewer of them following detection. Language plays a critical role in categorization, as it allows to gain new categories indirectly with “hearsay”, without the necessity to engage in the timely process of direct trial-and-error learning. I found this particular section of the reading interesting, as the capacity to efficiently and rather quickly transmit explicit knowledge without trial-and-error experience is what provides language the adaptive advantage unique to human beings. This made me think of other species and the form of categorization that they are capable of producing. Surely, there is great dissimilarities with the categorization made by humans but I suspect that since animals do not possess the same kind of language that we as humans use on a daily basis, they must therefore rely on trial-and-error learning, dependent on their specific sensory and cognitive abilities.
ReplyDeleteMelika, good summary. Now read the other replies (including last weeks'). Language is propositional: Why don't other species have it? And what, exactly, do they lack?
DeleteI think that many reasons could explain why and how we developed language and not other animal species, even though they can rely on other forms of communications, such as the Vervet monkeys who’s calls will vary depending on the threat, or the bee waggle dance. For the anatomical explanation, our larynx is further down our throat, which allows us to produce a wider variety of sounds and i think that some specific brain areas have developed specifically in response to our language abilities, such as the Wernicke or Broca’s areas involved in the production and comprehension of language. Another reason i was about to state is the selective pressure of the environment and the necessity for humans to cooperate in order to maximize the survival of the specie, but since many mammals live in groups without using a language as complex as ours, i don’t think this is sufficient to explain why other animals didn’t develop a similar method of communication.
DeleteBefore knowledge could be passed down through hearsay, the concept of learning through trial and error must have been an essential factor for survival. For example, how did the first humans distinguish between edible mushrooms and inedible ones, when foraging for food? Without features innately associated with unpleasantness for our senses (spiky texture, foul smell), it would be difficult to infer which mushroom to stay away from. Therefore, we can argue that learning to categorize edible from inedible mushrooms may have happened through the observation of others’ trial-and-errors. For example, by observing a peer eat a red mushroom and get sick, and another eat a brown mushroom and stay healthy.
ReplyDeleteThis point gives me pause, because if all non-innate categories are eventually based on supervised learning through sensorimotor experience, does that mean that categories that aren't vital to survival (like the poisonous mushrooms) or don't have a natural feedback mechanism associated with placing them in one category or another (like a fruit that smells or tastes bad) didn't exist before language? Say there's a caveman who hasn't developed language yet. Does he have a mental category corresponding to "mushrooms" in general, encompassing both the edible and inedible kind? Apparently he can't, because there's no mechanism to teach him how to categorize things as mushrooms. But that doesn't make intuitive sense to me - I feel like he'd know about mushrooms anyway.
DeleteAnais and Aya,
DeleteNonhuman animals as well as pre-language humans could already learn, and learn categories, long before humans evolved language. All the other means of learning were already available: unsupervised learning (classical/Pavlovian) and supervised (trial/error/reinforcement). (Imitation too, but that's not much help in learning categories: why not?)
The mushroom island is the model for supervised (reinforcement) learning. The "supervisor" is the inborn effect of eating something harmful: you get sick. That is the error-correcting feedback signal: negative (-) i you get sick, positive (+) if you get nourished. The +/- feedback is what guides your feature-learning mechanism (perhaps neural nets) in detecting and abstracting the sensorimotor features (striped stem, red cap) that distinguish the edible mushrooms from the poisonous ones.
There were no others to watch, on the island, to see what they eat and whether it makes them sick. And imitation is not really a means of learning categories; it's more for learning motor skills.
Language is a (very special) form of communication. It makes it possible to learn new categories from those who have already learned their distinguishing features: They can >tell you what they are. But how does this work? You know the categories that you have learned directly, through supervised sensorimotor learning. You can detect their distinguishing features.
What else is needed for you to be able to tell the distinguishing features of a category to someone who has not yet learned it?
Remember that all you have so far, like all other species, is the capacity for direct sensorimotor category learning by trial and error.
Exactly what more do you need?
In determining what else you need to be able to tell someone else about a category, I think you would need to understand what other categories that person knows or at least what other features they have learned to distinguish. This reminds me of the minimal grounding set that we discussed last week, in that you would need to use the features and categories that they already learned to explain this new category. If you didn’t know what categories they had already been exposed to, you would need to be able to give them the sensorimotor experience to develop the categories in the same way that you did.
DeleteTo learning categories through language, both people involved need to understand the context of the learning. For instance with the island example, if one person is to teach another about mushrooms, they need to understand the intentions of the second person, which they could potentially infer from the context: if the intention is to create a plant collection, the teacher may categorize white mushrooms as pretty and therefore pickable, but in the context of looking for food, the teacher should instead categorize them as poisonous and thus avoid. The teacher needs to be clear about what type of category they are describing (for example sensorimotor so that they know picking and eating the mushroom is a possibility), as well as describe the new category based on a shared pre existing knowledge of the features' categories.
DeleteAs language allows us to gain new categories without directly doing the hard work of trial and error ourselves, we can learn a much larger amount of categories. But we must also then learn which indirect sources to trust and not to trust. This may be more related to the hard problem of why we feel things when we are doing things or learning to do things; for example positive feelings towards a person may make us trust them more and thus be more likely to accept category labels they tell us. But if someone we highly dislike tells us a white mushroom is edible, we may not categorize white mushrooms into “eat” sensorimotor categories because we question that person’s motive.
I think it's very true that in order to *tell* someone the distinguishing features of a category, you need both context and and to share some categories between you, a sort of shared minimal grounding set, as Megan and Rebecah mentioned above. But for the actual act of telling someone about a particular category, there are even more basic things you would need. I'm thinking of propositional thought--the thing that gives you the little voice in your head or lets you have "verbal" but non-vocal thoughts. But propositional thought is so tied up with language that I'm not sure you could have one without the other, so maybe the criteria we're actually looking for to tell someone category features is something that allows humans specifically to have propositional thoughts. In that case, I would say language is what enables propositional thought, but then we're stuck in a bit of a loop! So is it possible to tell someone category features without propositional thought?
DeleteI found this paper’s discussion of the balance between generalization and uniqueness particularly interesting. While the paper takes the extreme examples of individuals with super memory, who recognize uniqueness to the exclusion of generalizability, I wonder how this can be applied across individual differences. I am imagining the tendency to generalize vs appreciate individual differences along a normal distribution, with most people sitting somewhere in the middle, at a point at which differentiation between grossly similar things is easy, but recognizing the commonalities is as well. Usually, in these ‘normal range’ people, we tend to see very few differences in the way things are categorized; there is a common language of categorization that is only really challenged at edge cases. I wonder how much of this fact is due to there being a socially and culturally common vernacular, which guides the ‘learning’ portion of developing categorizations, also mentioned in this paper. If so, and if we are to believe that a cultural understanding of how to categorize guides the way people actually categorize, then this would imply a. that the ability to do this categorization is in part learned and b. varies across cultures. I think this has interesting implications for how we communicate with people from different backgrounds, as well as what strategies we might use with AI when teaching it the ‘appropriate’ categorization strategy.
ReplyDeleteMy comment ties a little with what you said :
DeleteWhile reading To Cognize is to Categorize, very quickly a thought popped into my head about learned categories and culture. Though the text quickly addressed it, it was worth more attention in my opinion. Culture and experience obviously impact how we categorize things. Depending on where we come from, our categories change and even some members of those switch from one another. However, some items are in the same categories universally, more and more because of globalization but taking out the items in consequence of the latter it seems like the more the item is grounded the more it is universally categorized. Would this be a good way to approach cognition through categorization trying to get to the most universal level of categorizing to find similarities in different cultures – hence similarities in cognition?
Madeleine and Garance, what is a category and categorization? Before worrying about generality and culture, how do we get to the third, indirect way of learning categories (Direct: (1) unsup, (2) sup; Indirect: (3) words)?
DeleteIf I understood correctly, the three main ways we put things into categories are (1) Unsupervised Learning, where we group things naturally by seeing what's similar about them. (2) Supervised Learning, where someone teaches us, like when a kid learns to call all similar animals 'dogs' after seeing one dog picture. (3) Indirect Learning via Words, where we learn about ideas through talking and reading, not just by seeing or touching. These ways show how our own thinking, learning from others, and language all help us understand and categorize the world.
DeleteLast week, we were told in class that association is a weasel word, and to think of the hierarchical organization of the brain as some higher-order capacity to associate concepts with one another and lead to categorization of indirect concepts (my example was “love”) was fundamentally flawed. I then raised the question of whether we could replace “association” with “abstraction” and have the above argument be valid, to which the answer was no. After doing this reading, I wanted to clarify what I meant by linking the two concepts together:
ReplyDeleteWe know that to abstract (in the simplest terms) is to single out some subset of sensory input while ignoring the rest, and this is necessary for us to RECOGNIZE any kind of thing. Further, as seen in the examples of Funes and S, an inability to abstract and generalize to “recognize sameness, or similarity, or identity” hinders our ability to categorize (hence cognize, to a large degree). Now, while I agree with all this, we say that associating is a weasel word, but is it not just a part of recognizing?
I do digress that I was somewhat off in my wording: I do not mean to argue that “association” is synonymous with “abstraction”. To abstract requires us to selectively forget, or ignore, certain features, whereas association (i.e, Hebbian learning) is to link two concepts together based on some relationship between them or their co-occurence. By saying this I am challenging the weasel-ness of the word “Association”, but I am mostly wondering whether we can think of association as a necessity for recognition, or even less, a subordinate component of abstraction?
Paniz, unfortunately, "concepts" is a weasel-word too. Even if we substitute "category" (as defined in this course [how?]) for "concept", we have no idea what is being "associated" with what, and how: the name of the category? the feature-detectors of the category? the referent of the category? the members of the category? the sensory projections of the members of the category?
DeleteAnd associated how? As in word "associations"?, "Say the first word that comes to mind when I say 'chair'"? Or as in paired association learning with words or nonsense syllables? "Whenever I say "Gex" remember to say "Zof"?
It is astonishing how much we hear in psychology by way of explanation that really doesn't explain anything.
If there's something you have to learn to DO with the members of a sensorimotor category, like an edible mushroom (eat it) and not-do with the non-members (poisonous mushrooms) and we learn (e.g. with deep-learning nets receiving the optical projections of the mushrooms on its sensory surface) a convolutional network can learn, by trial, error, and error-correction "back-propagation", to detect the features that reliably distinguish the edible mushrooms from the inedible ones, so that the robot or body that contains the net becomes able to DO what what needs to be done with what. That is called "abstracting" the (sensorimotor) features of the category.
Does this help?
Could someone please clarify what input 'affordances' are?
ReplyDeleteHi Fiona,
DeleteAffordances are invariant sensorimotor features; they are invariable across situations and POV. It's basically a thing that our sensorimotor can recognize because it is invariant. He gave a boomerang as an example; we can recognize (i.e. categorize) a boomerang when it's static but also when it's in the air because it's shape remains the same. Unsupervised networks could theoretically come up with categories themselves by identifying affordances.
I see it as a fancy/elaborate version of geons; you can recognize the thing in all POVs because there is a certain invariance to it and we are thus able to deduce what it is supposed to be (ex: affordance leads us to hesitate between floor vs wall and our abstraction of other critical features allow us to distinguish between the two (in this case, the fact that former is horizontal while the latter is vertical)).
Let me know if that makes sense!
To my understanding, affordances can be defined as what an object "allows" or "offers" an organism to do with it, based on that object's shape and the organism's capacities. Categorizing an object depends on its shape, and thus its sensory and motor features, but also on how the organism can interact with it, in other words, the organism's capacity to detect these features in different contexts. This concept aligns with Chomsky's notion of 'competence' we discussed in class - what do organisms need to have inside of them that enable specific actions (although that mainly referred to linguistic capacity).
DeleteConsider a doorknob: for humans, equipped with opposable thumbs, it provides the means to open a door by turning. In contrast, for animals like dogs, a doorknob doesn't present the same possibility. Similarly, the ocean allows dolphins to utilize echolocation through sound waves, a capability not afforded to humans, even though certain sensory-motor attributes of the ocean, like its wetness or water transparency, remain constant.
However, trying to come up with different examples made wonder: aren't all sensory attributes essentially affordances? If we lacked the photoreceptors to perceive a particular color or the mechanoreceptors to feel a texture, such as a doorknob's metal, wouldn't that also alter our interaction with that object?
DeleteNatasha, you got it right: an "affordance" is a sensorimotor feature that allows you to DO the right thing with the right kind of thing. The emphasis is on DOing.
DeleteThe notion was JJ Gibson's, and it's become a bit of a cult. You miss nothing in this course if you just take it to be any sensorimotor feature that enables you to do the right thing wwith the right kind of thing.
Aashiha, that does make sense thanks! When I went back to the reading I saw the example of chairs and how a chair affords sitting-upon, so it couldn't really be a chair without it having the ability for people to sit on it.
DeleteIn relation to Natasha's comment, the doorknob has features which allow for people to turn it. This is only a salient feature for those that wish to use the doorknob in this way. Similarly, the chair has sensorimotor features which allow people to sit on it.
Fiona, the idea behind the doorknob is that it lets you in if you manipulate it the way that affords opening, not with the way you want to use it!
DeleteI found the criticism of Fodor's nativism interesting in this reading. It does seem unlikely that we are born with mostly innate categories, instead it seems more like our feature-detectors are innate which allow us to learn categories. I remember in my child psychology class learning that infants learn basic categories (tree) before subordinate categories (oak tree) and superordinate categories (vegetation), which are mostly taught by parents building off of the child's understanding of these basic categories. This might indicate that basic categories are ones that can be learned through more basic trial and error reinforcement or unsupervised learning from infants seeing patterns in features, which then provide the building blocks for more complex forms of abstraction that allow us to make categories of categories and understand context in categorization.
ReplyDeleteAdrienne, neither most of our categories nor most of our feature-detectors are innate: What's innate is our capacity to learn to detect new categories' features -- plus whatever default feature-detector settings we are born with: But the actual feature-detectors for all the new categories we learn, lifelong, are learned, as we learn the category. The learning mostly takes the form of detecting which of the many features of the members of a new category are relevant to distinguishing them from the nonmembers (which are usually members of other categories), increasing the weightings of those relevant features, and ignoring or down-weighting the features irrelevant to the categorization. The up-weighting makes the distinguishing features more salient, and helps make the category "pop out."
DeleteThe "basic level", "subordinate level" and "superordinate level" distinction is approximate. Categories are not strictly hierarchical. Their structure is more like a network, and can be represented by directed graphs
In the reading, it is concluded that cognition is categorization, but as mentioned in some of the above comments, it seems that it is more accurate to say, "cognition is mostly categorization"; that the process of categorization is "necessary to help us solve the easy problem", but that it does not cover cognitive actions such as continuous motor skills. I was wondering what role categorization plays in cognitive functions such as feeling, which remain unexplained and perhaps unexplainable (as per HP). As an example, we have the capacity to feel pain, and we can learn to categorize the felt experience of pain as being "pain". We can use the category "painful experience" to describe many different kinds of pain. My question is, must we be able to categorize pain in order to feel it? In other words, if we lacked the capacity to categorize, could we still feel? Intuitively, I would think we still could feel. But I don't know what such feeling would look like without the basic ability to categorize any part of the experience.
ReplyDeleteHi Adam,
DeleteI think we use the language “to feel” to describe / categorize a SENSORY experience. Thus, most of us are unconsciously categorizing this felt sensation through direct sensorimotor experiential learning and subsequently using language to describe it. Thus, we are learning about the category of pain through experience, as you said. My interpretation of your question is: if a person is unable to categorize using language, what is their awareness of the sensory painful experience? Here I think the person is still conscious of the painful experience, even if they lack the ability to selectively abstract the important features to describe it in words. What if a person lacked the nociceptive pain receptors to perceive pain? In this case, although the person may never have the sensory experience of pain, they still cognitively may learn to categorize what pain is through the descriptions of other people. I have never experienced childbirth but I cognitive can conceive how painful it is based on the descriptions of other people. Thus, my understanding of the reading is that categorization occurs in an “either or” fashion: either we learn the categories indirectly through experience OR we learn using language. Regardless of how the categories are formed, by interacting with our world we are still cognizing and thus, categorizing.
Adam, the capacity to categorize is the capacity to DO things.
DeleteDOing capacity (behavior, performance capacity) is observable and measurable and testable (with the T-Tests, T2-T4). And there's no reason to doubt that it can be reverse-engineered and its causal mechanism can be fully explained. (And Turing says Cogsci can't do any better than that,)
FEELing is not observable (except by the feeler). It's real, and we each know it, because we each feel.
Feeling is not T-Testable, though, so it's not part of the "Easy Problem," yet explaining feeling is still a problem for Cogsci (who else?). The problem is not categorizing feeling, whether sensations or emotions or pain or what it feels like to think, understand or mean anything at all. Categorizing is something we DO. The Hard Problem is explaining how and why feelings are being FELT at all, whilst we go about DOing the things we can do.
Kristi, why would anyone think that only talking humans can FEEL? (Sometimes I think it's the opposite, given the feelingless things humans DO.) It's as close to certain that most other animals feel as it is that other humans feel. The OMP is not the Hard Problem: Explaining how and why feelers feel is.
The following words in your comment are weasel-words: experience, unconsciously, awareness, conscious. They all refer to the same thing: feeling, sometimes redundantly (why say it 2, 3 times in the same phrase?) and some times in flagrant contradiction, unmasked by the de-weaselling.
Here's your original weaselled version:
“I think we use the language “to feel” to describe/categorize a sensory experience. Thus, most of us are unconsciously categorizing this felt sensation through direct sensorimotor experiential learning and subsequently using language to describe it. Thus, we are learning about the category of pain through experience, as you said. My interpretation of your question is: if a person is unable to categorize using language, what is their awareness of the sensory painful experience? Here I think the person is still conscious of the painful experience, even if they lack the ability to selectively abstract the important features to describe it in words.”
And now here is the de-weaselled version (I strongly recommend everyone does this same de-weaseling exercise with their own weasel-ridden passages until you wean yourselves from the practice!):
“I think we use the language “to feel” to describe/categorize a sensory feeling. Thus, most of us are unfeelingly categorizing this felt sensation through direct sensorimotor felt learning and subsequently using language to describe it. Thus, we are learning about the category of pain through feeling, as you said. My interpretation of your question is: if a person is unable to categorize using language, what is their feeling of the sensory painful feeling? Here I think the person is still feeling the painful feeling, even if they lack the ability to selectively abstract the important features to describe it in words.”
I think what you meant (or ought to have meant) to say was:
(1) You don't need language to feel.
(2) You don't need language to categorize what you feel. (All you need DO for pain is "ouch!" and run.)
Hi Prof, that was indeed what I was trying to say. Hard to kick the weasel word habit!
DeleteIn a class a while back we learned about prosopagnosia and how those with prosopagnosia recognize people or things by very selectively attending to certain features. For example a mustache would hold a lot of weight in identifying a person if that was a distinctive feature of that person, instead of their face as a whole. I find this disorder very interesting when thinking about how we weigh features more heavily when categorizing them. I’m assuming this kind of categorization (identifying people) fits into unsupervised models because the ways in which people look different are often quite easily identifiable and so being able to identify people just depends on exposure to them.
ReplyDeleteFiona, good points, though faces (and facial expressions, and voices, and speech) seem to be special categories to which a lot of evolutionary "preparation" has been given. (But face identification still requires some supervised learning, even if you don't have prosopagnosia.)
DeleteI wanted to begin my skywriting on the “To Cognize is to Categorize” paper by talking about the innate categories section. Although I have a hard time fully accepting innate ideas (the extreme form of nativism described in the paper where evolution doesn’t seem to matter), as I believe that any category we form must come from somewhere, I was wondering if Immanuel Kant’s ideas on a priori knowledge would fit into this section. Is it similar to Chomsky’s idea of universal grammar or is it something completely different and to be somewhat ignored like Fodor’s big bang idea.
ReplyDeleteIn addition, this paper in explaining categorization (which is doing the right thing with the right kind of thing as shown in previous replies) has helped cleared up some of the initial misconceptions I had with the concepts of supervised and unsupervised learning in the previous weeks. In relation to supervised learning, through trial and error, eventually, you will get to the level of abstraction that is required to have the category grounded. This idea helped me put into context last week’s readings on the symbol grounding problem.
Ethan, good reply.
DeleteYou don't have to worry about innate "ideas," because "ideas" is a weasel-word. If it just means categories, and categories are just kinds of things you DO certain things with (like eating edible mushrooms and avoiding poison ones), then, yes, organisms (including us) do have some innate categories: we run away from snakes and spiders at first sight; and we see the qualitative bands of the rainbow. No prior learning necessary, just a few million years of evolution to selectively shape the feature detectors and the action patterns. But most of our categories are not like that. They're learned.
Chomsky's another matter. Ordinary Grammar (OG) is a learned and learnable category. Chomsky's Universal Grammar (UG) is not (Chapter 8 and 9), because of the "poverty of the stimulus": To learn a (nontrivial) category, you need to have both positive and negative examples (members and non-members, with corrective feedback when you do the right or wrong thing). For UG there are only positive examples because no one -- child or adult -- violates UG. (There are plenty of violations of OG.)
Section 19 focuses on feature selection and weighting, examining Watanabe’s “Ugly Duckling Theorem” which proposes that our visual system selectively weighs certain features more heavily than others, making some features appear more salient. Using the story of the ugly duckling, he argues that the swanlet has no real reason for being marked as different from the ducklings, as its color is one of the only features that sets it apart, while the other birds differ on multiple features. If all features were weighted similarly, it would be hard for us to discriminate between objects. Fortunately for us, features are weighted, with color and shape being some of the most salient features. I was wondering if this effect is why the Stroop task is so difficult. In this task, participants are asked to name the color a word was written in. This is complicated by the fact that the words are themselves colors, but spell a different color than the shade they’re written in. Is the salience of color in our visual system responsible for this difficulty, or are shape and color similarly weighted, resulting in the confusion? The reading lists shape and color as two of the more salient features for our sensorimotor systems, so my thinking was that the disparate shapes and colors presented in the task are “battling” importance.
ReplyDeleteThe idea of feature selection and weighting in the context of the "Ugly Duckling Theorem" is indeed an interesting concept when considering tasks like the Stroop task. I previously learned in a cognition class that the stroop effect is produced through interference, where part of your processing occurs automatically. The reason for this difficulty can be linked to the salience of features in our visual system. Color and shape are typically among the more salient features, and as you mentioned, they can "battle" for importance when they are in conflict. So in the case of the Stroop task, color and word meaning are both salient features, and when they conflict, it creates cognitive interference. So yes, like you said, I believe that this battle for importance can be related to the idea that our visual system selectively weighs certain features more heavily as described in the ugly duckling theorem.
DeleteThese are interesting points! This also made me think of the oblique effect. The oblique effect is when observers are able to discriminate between lines that are vertically and horizontally oriented more quickly and accurately than those with an oblique orientation. In a previous class, I learned that the oblique effect could be due to experience-dependent plasticity, where the response of specific neurons can be shaped by our perceptions of the world, like the fact that our environment is made up mostly of horizontal and vertical lines compared to oblique lines, which is why these features are weighted more heavily. Could this be related to learned CP?
DeleteThe ability for us to communicate unique abstractions through language is so interesting and I'd never thought about it before. What is the mechanism that allows us to understand (in the felt way) the application of two features that I know but have not been important to me in combination before and then turn that into a category? I.e. avoid eating long red-capped mushrooms. It feels like something to see an instance of a category that has been described to you. Recognizing something you've never seen before seems incredible now that its been explained to me in those terms.
ReplyDeleteIf I’m understanding your question correctly, you are curious about how we come to feel things when we combine familiar features in new ways to generate a category - and how language helps us in the process of understanding abstractions. I’d suggest that the answer is simply category learning.
DeleteLast class, “Stevan said” something that really stuck with me. Ideas that are feeling-related (e.g, love), subjective stuff (e.g, warm, red) and even the feeling of understanding can be placed in a felt state category (not some meta-thing out of this world). Noticing concrete occurrences of abstract features that were described to you feels incredible indeed. I’m not sure you can “understand [something] in a felt way” but you can associate a feeling, a qualitative state with the grounding of symbols. From the reading I also got that language acts as some kind of shortcut in our learning and categorization such that we don’t have to go through the process of sensorimotor learning, but we can understand relatively quickly through hearsay.
The concept of feature weighting mentioned in 19 really stuck with me. It says that features are arbitrarily selected and given extra weight. But wouldn't feature selection also be subjectively weighted as well? I am not sure what the arbitrariness is due to but even if it is just the way the visual system functions (like contrasts, continuity, those Gestalt things), I have a hard time believing everyone perceives features the same way. How could individual experience not affect the weight of certain features? It is possible I'm going to far and all we're saying is that the visual system detects some things more than others.
ReplyDeleteI was also wondering the same thing when reading this section. It seems to me that since categories and feature detectors are not innate (to my understanding) that it would be impossible for personal experience not to play a role in determining what features are most salient. Various upbringings and culture would in this way affect what features we are able to most strongly perceive. Unless this category of salient features was innate in humans, I don’t see how we would all be able to learn to perceive the specific features consistently given various social factors.
DeleteThe paper mentions how our categories are based on the ways we behave towards different things in the world, and that and our language plays a role in shaping those categories. This could raise the question of how categorization processes differ across cultures and languages, This question has important implications for cross-cultural communication and understanding, as well as for fields like psychology and neuroscience that rely on categorization as a fundamental concept.
ReplyDeleteHi! I think that's an interesting point. It also relates to the reading "The great Eskimo vocabulary hoax", where Pullum explains that it is a myth that eskimos have different vocabularies to describe different types of now. This myth has often been used to illustrate the idea that language shapes perception and cognition. So maybe in the end we do not categorise that much differently from other cultures?
DeleteJessica, I had similar thoughts while reading this week’s piece. One thing that it made me wonder about was whether it is possible that rather than our language shaping the features we abstract, the features which certain groups have a larger focus on are abstracted in finer detail and language reflects that. To explain what I mean, imagine two countries 1 and 2 which speak languages A and B respectively. In country 1, there is a huge cheese industry and the production of cheese is a very large part of the country's culture. Consequently, there are an array of terms in language A which are used to convey incredibly precise details about different cheeses. In country 2, however, cheese is just another food, and other than a few categories such as “mozzarella” or “swiss”, language B does not include much cheese vernacular. If two people from each country were asked to describe the same cheese, the person who comes from country 1 and speaks language A is likely to abstract features of the cheese with a high degree of precision and convey far more detail than the other. This is not due to their language though, but rather what the language reflects about what is important to those who speak it.
ReplyDeleteAbstraction, as discussed in this reading, is a crucial ability of human beings. Not only does it allow us to categorize, but it also allows us to negate non-essential details in favour of more important ones which may be useful in pursuit of a goal. An example that this reading reminded me of is the famous selective attention test video in which a group of people are dribbling and passing around a basketball. The viewer is prompted to count how many times the basketball is passed around, and after the video is played it is revealed that a person in a gorilla suit had crossed the scene. This comes as a surprise to viewers, to whom this gorilla-person went unnoticed. Our ability to negate unimportant information allows us to focus on certain information in pursuit of a goal, such as counting passes. While we may not realize it, we are engaging in abstraction constantly, for the things we choose to listen to to the human features we implicitly focus on when recognizing individuals. As discussed in the reading, an individual such as “S” with a far more powerful rote memory than the average person would likely struggle to complete tasks which required selective attention and abstraction.
ReplyDeleteI still remember watching that video with a friend - I was so focused on keeping up with the people in white shirts and counting how many times they passed the ball, I did notice a dark figure going in the middle, but my mind just blended it in with the others in black shirts and chose to ignore it while solely focusing on the people in white, so I didn’t look at it long enough to notice it was a ‘gorilla’. My friend on the other hand was caught off guard by the sudden entrance of the gorilla, that they lost count of the number of times they passed the ball. It’s so interesting to see how important abstraction is within our lives, and how we all have different degrees of this ability.
DeleteBased on the reading, I think the key method through learned CP is through JND. It tells us what is the smallest range to tell a difference between items, and by increasing the number of JND, this may lead to a categorical difference in distinguishing two objects. However, people have different sensitivity, in other words, the size of JND varies from person to person. How we can possibly set the boundaries (or what is the minimum number of JND) to be able to make a categorical separation? It is really subjective. These concerns make me worried that is "hearsay" accurate enough to have new categories for people instead of direct trial-and-error learning?
ReplyDeleteThe paper summarizes in clear language the relationship between cognition and categorization in the context of sensorimotor systems. It emphasizes that organisms, as sensorimotor systems, interact with the world based on what their sensorimotor contact "affords." Categorization is described as a systematic, differential interaction between an autonomous, adaptive sensorimotor system and its environment. It involves the ability to detect and respond to specific features or invariants in sensory input, allowing for the differentiation of input types and the assignment of those inputs to specific categories or kinds. A strong argument is formed that categorization is a fundamental aspect of cognition, as it is about how organisms behave differently toward various stimuli and objects, encompassing aspects such as the detection of affordances, invariants, and the processing of sensory input over time.
ReplyDeleteThe paper argues that cognition, essentially, is categorization. Sensorimotor systems interact with the world based on what these interactions "afford;" the idea then for the system is that you can sense the world around you and interact with it based on what you can do with it. These affordances depend on these systems for categorization.
ReplyDeleteCategorization involves systematic, differentiating interactions between an adaptive sensorimotor system and its surroundings, allowing it to respond differently to various inputs. While Fodor and some suggest that all categories are innate, evidence from the paper indicates that most are learned through trial and error, with learning closely related to adaptive changes over time.
In relation, I want to discuss the mushroom example from class. Can a kid tell if a mushroom is safe to eat or not just by looking at it? Probably not, unless they've learned which ones are safe.
DeleteIn supervised learning–able to guide categorization, it's like having a guide on the island who tells you which mushrooms are safe (positive feedback) and which are not (negative feedback).
Then, with unsupervised learning, it should simply be more like wandering around the island with no guidance. So, there is a lack of how to categorize mushrooms. Thus, language is quintessential here to describe them, allowing understanding and communication for others as well.
Hi everyone!
ReplyDeleteI really enjoyed this reading! Stevan Harnad's article on categorization and cognition argues that all cognition is based on categorization. He suggests that our sensorimotor systems play a crucial role in how we categorize things, and that language also plays a huge role in this process. The article also explores the differences between explicit and implicit learning in relation to categorization. I thought the article was very thought-provoking and it made me wonder whether there are any situations in which categorization may hinder our ability to perceive and interact with the world around us, rather than enhance it. It seems that categorization may have inherent limitations but let me know what you think!
In the reading to cognize is to categorize, the idea of categorization being at the core of human cognition is brought up. I was thinking about the implications of this idea in the real world. If it is in fact true that our ability to categorize is innate, what are the implications of that in terms of the education systems found around the world? Does this in turn mean that the differing ways of teaching found from education institution to education institution in terms of categorization (specifically at a lower level of education) does not matter as much as some think it does? Moreover, could there perhaps be more effective ways of teaching younger generations through updated curriculums that incorporate this concept of categorization being a fundamental aspect of learning to adhere better to children's learning. As opposed to commonly found ‘just memorize the material for the test’ in school classrooms.
ReplyDeleteHi Stefan,
DeleteI agree with the points you've raised. Routine memorization and standardized testing have traditionally been prioritized in many educational systems, frequently at the expense of the basic cognitive processes that support real learning. Educators may create curriculum that actively encourage students to categorize and conceptualize knowledge by recognizing the value of categorization in learning. I think this will definitely help students develop a better comprehension of the material and their ability to think critically.
I like this response from Stefan Vujicic. It raises an interesting point regarding the implications of categorization being at the core of human cognition for the educational system. If educational institutions recognize that categorization is innate, that children naturally possess this ability, schools could emphasize activities that harness students' categorization skills that include activities encouraging them to categorize information, make connections, and draw relationships between concepts (rather than solely focusing on rote memorization). Categorization involves recognizing connections between seemingly unrelated concepts. Educational systems can promote interdisciplinary learning to help students categorize information across various subjects. This approach would go beyond memorization by allowing a deeper understanding of the subject matter while also supporting the development of well-rounded individuals. Consequently, it would force schools can adopt more adaptive teaching methods (visual, auditory, movement, etc.), helping students grasp and categorize information more effectively.
DeleteHi Stefan! I really liked your approach of connecting our innate ability of categorization with the predominant education system majorly focusing on rote memorization. Personally, I think integrating both learning styles can help students learn better. Especially at lower levels of education when they don't have any basic background knowledge of the subject, I think simply learning the material by understanding the concept and memorizing basic parts are essential (ex. multiplication charts). Afterwards, when learning more advanced higher level material, learning styles that promote categorization can be used (ex. instructors can do a flipped classroom approach, where the student can learn the material by themselves first, and do active learning (incorporating categorization style learning) in class).
DeleteI never realized the important role categorization plays on how we think and perceive things. Although we have some innate feature-detectors and the ability to recognize different features used for categorization, categorization mostly relies on affordance and learning. Abstraction goes hand in hand with categorization in our perception as it determines what specific feature we focus on and therefore how we react to it.
ReplyDelete
ReplyDeleteThis sentence in the article impressed me deeply: “mere cognitive scientists (sensorimotor roboticists, really) should not presume to do ontology at all, or should at least restrict their ontic claims to their own variables and terms of art -- in this case, sensorimotor systems and their inputs and outputs” (Harnad 22). Cognitive science does not necessarily have its own definite rules to define a certain thing.
There are many things in the world that are not variables or with data values. It does not have the process of programming and obtaining a series of output through input.
For example, when faced with a cup of coffee and a cup, what appears in the cup is a cup of coffee. This is something that can be touched, and they are touchable entities in the world, it has its own variables, and has its own values.
But when we encounter something that is intangible and invisible like beauty, like feeling, like truth. It does not have the properties of variables we mentioned before.
Another example is what we call God. Unlike this series of intangible entities, the variables of some entities of the spirit class are often different from the definitions of the tangible things we mentioned before. Therefore, cognitive science cannot only rely on one substantive data and variables to think and analyze problems.
Categorization is a capacity to learn about distinguishing features and disregarding irrelevant features, that requires an absolute discrimination capacity. The means to categorization are supervised learning (trial-and-error and corrective feedback until success), unsupervised learning (without corrective feedback, via the input shape), and indirect, language-based learning. Learning through language helps us to avoid the time-consuming process of trial-and-error learning and corrective feedback, just through hearsay. In other words, someone else tells you the features, without you having to interact (sensorimotor) with them directly. However, this indirect verbal learning is already grounded in sensorimotor categories, thus it begins from direct grounding. I was wondering about “recursive grounding”, in which Harnad (2005) writes “the words in their explicit descriptions must themselves be grounded, either directly, or recursively, in sensorimotor invariants”. I understand how it should start with direct sensorimotor grounding via our interactions, yet I was curious about how can it work recursively. Is it related to hierarchical categories? I was just wondering if anyone has any examples of what it might look like. Thank you!
ReplyDeleteI personally enjoyed reading the Recoding and Feature Selection passage. The passage explains that increasing dimensionality does not substantially improve categorization abilities. It highlights how transforming information into meaningful units, or recoding, can enhance memory capacity. It uses the example of converting binary digits into meaningful triplets to illustrate how this process can significantly improve memory performance. However, it also shows the challenges related to the credit assignment problem in machine learning, emphasizing the difficulty of determining the most effective features or rules from numerous options. In additon, it emphasizes the role of supervised categorization training.
ReplyDeleteHey Julide, I found the Recoding and Feature Selection passage quite intriguing as well. It underscores the importance of dimensionality reduction, while increasing dimensionality might not significantly enhance categorization abilities, the concept of recoding is fascinating. Transforming information into meaningful units can absolutely improve memory capacity. The example of converting binary digits into meaningful triplets that you provided is a very good example of this process. I wonder if the idea of doing something out of the ordinary during a mundane task, like knocking on your front door before locking it, to remember if you did it or not later is related to recoding? You also mentioned how supervised categorization training is a great reminder of the role guidance plays in machine learning, which I strongly believe emphasizes the importance of expert input in creating effective artificial intelligence models. Thanks for providing such a concise overview!
DeleteHi Zoe,
ReplyDeleteI find your comment to be really interesting with the example of country 1,2 and how the differing degree of significance of cheese in each country effects each languages A and B. I really agree with your comment and your comment reminded me of a real life example with languages. As each country has vastly different cultures, and each of the countries all have differing larger focus on different features, this is why there are some expressions that can be expressed in language A cannot be translated in language B. For example, in Korean, there is an expression called “Han-한”, which is often translated as sadness, regret, or resentment in English. But this fails to capture the full meaning of the expression. Also another word that does not have a direct English translation is “gosohada-고소하다”. This is often translated as nutty or savory in English, but it also fails to fully express the meaning.
I’m not clear on the underdetermination problem. Is it that all objects/things/categories (flower) have multiple features that overlap with other things and can fall into many other categories (plant, organism, colorful)? And because of this, as well as invariance and our general “poverty of stimulus”, it should be impossible to correctly learn categories? Harnad suggests in his text that supervised categories are even more underdetermined than unsupervised ones (13). Is this because we learn more ‘physical’ categories with unsupervised categories, which are constant and simple like depth or figure-ground perception? (an object is either close or far, a landscape has depth or does not, and that’s that) Compared to supervised categories, like flower, contains more features and is related to more categories.
ReplyDeleteHi Csenge,
DeleteMy understanding of the underdetermination problem is as follows:
1) An infinite amount of features can be attributed to a given object.
2) Therefore, in order to categorize an object in a certain manner, we must give more weight to certain ones of these features. (This is related to Watanabe's Ugly Duckling Theorem: given that every object has an infinite number of features (1), it follows that the cygnet differs from duckling A just as much as duckling A differs from duckling B. Thus, the fact that we distinguish the cygnet from the ducklings shows that we select and attribute more weight to certain ones of these infinite features (in this case, the feature of being "ugly.")
3) The underdetermination problem asks HOW we determine these features to which to attribute more weight to.
What you're mentioning - that "all objects/things/categories (flower) have multiple features that overlap with other things and can fall into many other categories (plant, organism, colorful)" - sounds more like the Vanishing Intersections problem (section 14 of the paper.)
Csenge, underdetermination and approximation are simple, and related: Things have a lot of features, and they can be members of lots of different categories (potentially an infinite number), based on different combinations of features. But only a few of the features are relevant for distinguishing the members from the nonmembers in any particular category (say, edible vs inedible mushrooms on an island). So you just need to be able to detect enough of them to categorize correctly in that context.
DeleteThis is true whether you are learning features by direct sensorimotor trial and error or you have been given a verbal rule (by a person who knows, or by a book) that describes the distinguishing features in words (on condition you already know each feature-names referent.
TFeatures are approximate, because you may eventually run into other mushrooms for which the features that have been enough to sort all the mushrooms you've seen so far, but now they're not enough and you need to detect more features. That tightens the approximation. But it's still underdetermined, because there's always a tomorrow, so you may need more features in your feature-detector or in your verbal rule.
Unsupervised learning is passive learning, picking up feature-feature correlations, but those features may not be the right ones when you have to DO the right thing with the members, and to distinguish them from the non-members. And that's necessarily supervised learning (why?). It's still a correlation, but between what and what?
There is a relation between underdetermination and the poverty of the stimulus (POS) for Universal Grammar (UG) (Chapt 8 & 9), but it's not very straightforward: I'll explain once you tell me what you mean by POS...
Ohrie, you're right, but we don't "attribute" features (at least in sensorimotor category learning): we detect and abstract them from direct experience, guided by the need to find the features that reliably distinguish the members from the nonmembers.
You're right about overlap, underdetermination and vanishing intersections too, but remember that the distinguishing features can be disjunctive (either/or): the edible mushrooms don't all have to be red or all have to be green: the distinguishing feature could be "red OR green". The invariant feature is neither red nor green but red-or-green. It's especially easy to describe such compound Boolean features verbally, as in Google searches, using not, and and .
But it's more intuitive to say "distinguishing feature" rather than "invariant" feature, because distinguishing category-members from non-members, and distinguishing categories from categories is what it's all about.
The reading "To Cognize is to Categorize" suggests that categorization must serve a purpose for the organism using it. This ties into the category-proxy distinction. We often want to know an object's unobservable traits but can only see its sensory shadow. Since sensory shadows are not fully reliable (Laplace's demon could duplicate any), probabilistic reasoning helps link proxies to categories.
ReplyDeleteIdentifying pets by their dietary needs is simple: pointy-eared animals (cats) need meat, while floppy-eared ones (dogs) don't. If these proxies, and others like them, stopped correlating with the meat-requiring distinction, the "cat/dog" category would lose its purpose for this task. Similarly, antiquity sea-dwellers classified whales as large wish, while modern biologists don't. Neither view is "right" or "wrong"; they simply used different categories symbolized by the same words, since they had different purposes for using these words. The terms overlap because core examples for each category (e.g., a tiger is not a fish, a salmon is) are consistent between the groups. Mixing up words and their intended categories is a mistake. Then, every object described by a word like "table" will fit into some plausible category within the range suggested by the set of dictionary definitions of "table", so the "vanishing intersections" problem (point 14 in the text) seems to be a conceptual confusion between some specific category implied by a dictionary definition, and the space of all reasonable categories thus implied.
Thomas, most of what you say here was not taught in this course. Nothing about proxies and core examples. Could you reformulate (and re-think) this in terms of unsupervised and supervised learning category-distinguishing features? (And see Appendix 1.)
DeleteThe thing that stuck out the most to me in this reading is the uniqueness of human hearsay due to our language ability. Which I understand to be the sped up version (verbal input) of somebody else’s long trial and error process (supervised learning). They learned the most “pop out” invariable features and can simply pass that along to another person, as long as the explanation is completely grounded in the learner. This special feature of language is a huge time saver and it logically makes sense why human technology grows exponentially. And yet we still waste time learning how to do manual long division in primary school instead of mastering the use of scientific calculators (joke, kind of).
ReplyDeleteKaitlin, good points, but it's not just speed and efficiency that language adds to imitation and unsupervised and supervised learning, What is it?
DeleteLanguage is a third way to learn categories that other species do not use. It allows the learner to learn a new category without having to interact with the category directly (sensorimotor abilities), through telling by the speaker. For the speaker to tell the learner a novel category, they would need to share a common symbol grounded language, which means all of the features used by the speaker have meaning to both the speaker and learner.
DeleteThe other-race-effect is an interesting example of how categorization is learnt. People recognize and distinguish faces of members of their own race better compared to members of other races. However, individuals can resolve this effect by being exposed to multiple races, which allows their visual system to detect and extract facial invariance, resulting in facial visual constancy.
ReplyDeleteI know!! I actually looked into this topic recently for a different class and I find it interesting what the neural mechanisms are behind this. So this is caused by something known as synaptic pruning, in which an elimination of synapses to increase the efficiency of neural communication occurs, which then would lead to the perceptual narrowing that we experience as we grow as adults! On the contrary, Synaptogenesis is the formation of synapses between neurons and this occurs at a very rapid rate when we are babies, just when we are born!! It's why babies who are 3 month olds can distinguish between faces of all races but 9 month olds can only distinguish between faces of their own race. I wonder if we can apply the concepts of synaptic pruning and synaptogenesis and generate a process that imitates them, how would we apply them to AI?
DeleteMiriam & Mallik, good points. How are they related to (1) underdetermination and approximation and to (2) context-dependence of categorization "compared to what?")? (I'm not sure "pruning" fits face-recognition as well as it fits phoneme-recognition. After all, we produce speech but not most of our facial features. It may fit the recognition of facial expression better, because that too is a mirror-capacity.)
DeleteBased on what I understand, when we apply underdetermination and approximation to distinguishing the races of faces, there are features that a variety the features that we perceive on other people's faces can allow them to be members of lots of different categories (in this case race), based on different combinations of features. That in itself is underdetermination, so we rely on approximation to distinguish the members from the nonmembers by focusing on the MOST relevant features to a specific race compared to other races. To address the synaptic pruning concept, I never thought of this that way actually, so I believe I misrepresented the relationship between our inputs of facial features and what we make of them, so in this case, to refine our categorization process, we just need to be exposed to more faces to create a bigger 'database' of features, which then would refine our capabilities to categorize).
DeleteI had never before thought about the fact that there is no innate category boundary between black and white the same way that there is an abrupt change in category between green and blue. I understand this, but I wonder whether it would be equivalent to simply say that when distinguishing between black and white we are actually choosing between one of three categories—black, gray and white—and when we are distinguishing between green or blue, it is the same thing, just with only two categories—green and blue. Just as the boundary between green and blue is arbtrary, the boundary between black and gray is arbitrary and the boundary between gray and white is arbitrary. To me, distinguishing either black from gray or gray from white is the same as distinguishing green from blue. Is it only true that there is not an innate category boundary between black and white by virtue that there IS an innate category boundary between black and gray and one between gray and white or am I missing something?
ReplyDeleteJordan, you're right, and if you think of it, since these are all continua, not discrete-feature based categories, it's to a great extent a matter of scale.
DeleteWhen I read about learned categories, an example came to mind. This kind of categorization is closely tied to learning in a very specific period. Any child younger than six who receives musical instruction, such as playing a piano with labelled keys or listening to adults match tones with words, might develop perfect pitch (the ability to identify a musical note correctly upon hearing it). It is tricky because once you pass this learning period, it becomes difficult to regain this ability. Although everyone has innate pitch feature detectors, not receiving exact learning during this critical time means you won't have this musical categorization perception again. This example is strong enough to illustrate the point that “categorization is intimately tied to learning.”
ReplyDeleteJinyu, yes, and those who develop perfect pitch may also develop a CP effect for pitch categorization and discrimination.
DeleteIn the “Abstraction and Amnesia” part of the reading, Professor Harnad states that, “Borges portrayed Funes as having difficulties in grasping abstractions, yet if he had really had the infinite memory and incapacity for selective forgetting that Borges ascribed to him, Funes should have been unable to speak at all, for our words all pick out categories bases on abstraction.” This part was so interesting for me to read, especially because I haven’t noticed that I was using abstraction all the time, while I was speaking and recognizing things. I thought that having a perfect memory would be a gift, since the person wouldn’t have to try memorizing, but would know everything with little effort, however learning that it can become a handicap was beyond unexpected. Moreover, a real life example of the stage memory artist “S”, and his inability to concentrate while reading novels, was shocking.
ReplyDeleteHi !
ReplyDeleteCould someone please explain me what is "hearsay"? Is it simply the use of language to categorize, or is it more complex than that?
Thank you in advance !
Juliette, it's the same thing as « ouï dire », but it's mostly just metaphor for language.
DeleteI found the idea that categorization is inherently tied to the ability to forget to be very illuminating, because I think it relates memory, abstraction, and categorization by associating them with the adaptive value of imagination. In order to form categories (eg. ducklings vs. cygnets) we have to selectively forget features of the world that are impressed on our sensory systems. If we are unable to do this, we would not be able to form categories, since the difference between a pair of ducklings is just as significant as the difference between a duckling and a cygnet. Furthermore, we would not be able to engage in imagination, because, as the case of the journalist S suggests, we can only recall specific instances of our lives, and are unable to use abstract categories to imagine novel situations, as is the case when reading a novel.
ReplyDeletePut simply, the ability to selectively forget certain features is a prerequisite for the formation of categories. The ability to abstract from our experience (through selective forgetting), then allows us to engage in imagination, which helps us generate novel internal models of external reality, which has clear adaptive value from the perspective of predictive processing models of cognition.
Daniel, there's no doubt it's useful to be able detect both similarities and differences. But kid-sib could not understand your last sentence.
DeleteAs explored in “To Cognize is to Categorize: Cognition is Categorization,” I particularly enjoyed relating categorization to different contexts (In the reading, this was referred to as "context-dependent categorization" (Harnad, 2005))--- for example, categorizing can be used to focus-in on the useful features of the world that promote survival. As mentioned in the reading, although some aspects of it may be innate, categorization is mainly and significantly related to learning (Harnad, 2005). It would be interesting to explore in further depth the advantages and disadvantages surrounding the different types of learning in varying situations. The idea that humans and non-human animals, as sensorimotor systems, can execute certain actions and then ameliorate our behavior from the outcomes of those actions, is said to be a form of categorizing that involves feedback-reliant supervised learning. On the other hand, unsupervised learning involves forming correlations.
ReplyDeleteIn terms of evolution and adaptive specializations observed in non-human animals, I believe supervised learning provides a greater advantage for survival. An organism acquiring an adaptive trait through trial-and-error (supervised learning) would establish cause-and-effect relationships, thereby leading to improvement and better navigation of their surroundings. However, the method of simply finding patterns in data (unsupervised learning) won’t be advantageous in this specific context. As Professor Harnad highlights in this reading, categorization can differ depending on the situation at hand. Learning how an object and its functions can be applied is key for survival— feedback would be required to identify these crucial features, as how something is used can vary depending on context. Simple correlations wouldn’t be effective as an object could have many uses that may not be relevant in those circumstances. I’m curious as to how these types of learning could be applied to other real-life examples— what are some more specific cases where context would change the effectiveness of these types of learning?
Michelle, you asked about:
Delete"context-dependent categorization"
Information is the reduction of uncertainty among alternatives that are important to you -- your success or your survival (like the (vegan) sandwich machine).
In supervised category learning, you are trying to reduce uncertainty about what to DO with what kind (category) of thing.
On the mushroom island, your survival depends on (or your feature-learning neural-net) detecting the features of the mushrooms that are edible to you. Your rescue may depend on igniting flares at night by lighting mushrooms as flares. But only some kinds of mushrooms can be lighted as flares. And their features may have nothing to do with whether they are edible or inedible.
Two ways to categorize the same mushrooms, depending on the context (food or rescue); two different sets of distinguishing features.
“unsupervised learning involves forming correlations”
unsup learning is based on learning to detect feature/feature (S/S) correlations (and frequency).
“supervised learning provides a greater advantage for survival”
Sup learning is based on feature/outcome (S/R) correlations (based on the consequences of what you DO with what kind (category) of thing.
(If the S/S correlations turn out to also be correlated with the S/R ones, you were lucky; if not, you have to work harder to find the winning features. The “luck” may be thanks to evolution, if those features were already important for your ancestors: Chapter 7.)
“how something is used can vary depending on context.”
The right thing to DO depends on the context of alternatives among which you have to learn the distinguishing features (through unsup, sup, or words) that will reduce the uncertainty about what to DO with what to as close an approximation as you need.
“what are some more specific cases where context would change the effectiveness of these types of learning?”
Think of more example like the edible mushrooms and the flare-able mushrooms in real life.
I don’t think I quite understand sections twenty and twenty-one, discussing discrimination vs. categorization and the impact of recoding on categorization. I understand Millers proposed limits on information processing in the brain, with the famous number of 7 +/- 2, but I don’t quite understand how this applies to categorization. I understand the distinction between discrimination and categorization, the former being relative discrimination (i.e., telling if two stimuli are same or different) and the latter being absolute discrimination (i.e., telling what a stimuli is). In Miller’s explanation, the limits on absolute discrimination are far narrower than those on relative discrimination, in that the differences have to be far greater for identification than discrimination.
ReplyDeleteHowever, while I understand all of this perfectly well, I’m having trouble with the claim that “the number of regions along the dimension for which we can categorize the object in isolation is approximately seven”. I understand this to mean that, if we are attempting to categorize an object based on a sensory dimension such as size, there are approximately seven different categories of size that we can reliably identify, and more importantly use for categorizing the object’s size (which is one of its features). Is this correct? If not, could someone take the quote and put it in more kid-sib terms?
I also understand the concept of chunking/recoding and how that can improve our information processing, but I don’t really understand how this applies to categorization. How does the recoding process work with respect to categorization, and how does it improve it…can someone explain this in terms of the paper’s argument that categorization is “the systematic differential interaction between any dynamic and adaptive sensorimotor creature and its environment” as well as the discussion in the paragraph above (re: seven regions along a dimension that can be used to categorize an object)? Is there an example of recoding being applied to categorization, similar to how you can recode the decimal system into the binary system?
Stevan, yes, your understanding of Miller's 7+/-2 limit and everything you said before it is correct.
DeleteSeven was just an empirically discovered limit on subdividing a sensory continuum and on short-term memory (which has since been revised and updated many times).
About chunking and recoding: Direct sensorimotor category learning is itself already recoding: DOing the right thing on the basis of selective sensorimotor feature-detectors. The DOing can be any differential action, not just naming.
But if a species has language (propositions), and its members have directly grounded enough category names (including the names of features -- round, red -- which are also categories), those category names can be used in propositions to describe the features of further categories.
So if speakers know which are the distinguishing features of a category "C," they can TELL them to hearers who already know those features' names, but don't yet know that they are the features of C.
That is all recoding/rechunking, exactly as in Miller's example of learning to increase your short-term binary digit span by memorizing the code for assigning them their decimal names.
(It's also the meaning of the mathematicians' slogan that "The test of a good theorem is whether it generates a new definition. On the mushroom island, if more people arrive, you can say 'the "edible" ones' instead of 'the "long-stemmed, red-capped, white-dotted" ones' -- and you can even put that into your dictionary.)
In this reading, I thought about the ability of an individual with akinetopsia to categorize objects and events involving motion. I think that these individuals must rely on unsupervised learning in which their visual perception is characterized by focusing on static, spatial elements and an emphasis on non-visual cues, akin to how unsupervised learning clusters data based on inherent features without external, labeled feedback. For example, a person with akinetopsia observes a fan, and since they lack the ability to perceive continuous motion and temporal patterns, they only see snapshots rather than a continuous rotation. This perception makes it difficult for them to assign a single, continuous label to the fan’s motion such as “spinning.” However, through unsupervised learning they can group snapshots with similar blade arrangements as “fan configurations,” and over time learn to create categories like “fan with blades arranged in a circle,” produces wind.
ReplyDeleteI wonder about UG being alone amongst the innate. I feel like there is evidence for other social things having a certain degree of innatism amongst the average human, we are very much evolved for social matters, and in particular facial recognition jumps to mind.
ReplyDeleteMy baby brother would look me in the eyes since fresh out of his mother, the fusiform face area is (normally) for well... faces! and by far the most common pareidolia is seeing faces.
Other things come to mind, cuteness for one seems like more than just hearsay, Kindchenschema can be argued to have quite a strong innate component.
Of course these are all still grounded sensorimotor-ly, but like language, maybe there can be other "very human" categories that are evolutionarily old enough to be innate.
I couldn't find the comments I wrote last week so I'm posting them again here, sorry if there are duplicates. The innate categories section made me think about the fascinating concept of inherent differentiable responses. It's intriguing to consider that we might be born with the capacity to differentiate between various categories without explicit learning. Jerry Fodor's perspective adds depth to this notion, suggesting that such innate capabilities exist. This leads me to question the extent of innateness in human cognition and the factors that influence individual variations. If we're all born with the potential to respond differentially to categories, why do our responses vary? What role does nature versus learning play in shaping these innate capacities, and how do they manifest uniquely in each individual?
ReplyDeleteIf I understand this paper correctly, to learn about something is to learn about how to categorize its features, from perhaps most salient to least salient. And this salience depends on what sensory systems are most weighted in our biology (for humans, sight. For dogs, smell...) Say an apple specialist will be able to recognize a Gala apple from a Cortland apple based on the size, taste, coloring, etc. The same way eunology enthusiasts can sometime recognize the soil in which the grapes were grown. All of this comes from trial and error categorization, comparison (discrimination) and testing. A not so novel thought is that we do the same thing with people(other sensorimotor systems) as we get to know them. Our knowledge of their (most niche) interests expands and never really stops expanding, because they too have infinite amounts of features to recognize, categorize and so on. A stranger will appear a certain way, and in microseconds we have already categorized them (for without categorization, the world would be so stupidly complex that we would not be able to parse anything from anything -- thinking about Funes here). I'm thinking about PSYC 331, where we learned that androgyny (or looking both masculine and feminine, to the point of categorical confusion) was frustrating to many. And this frustration was tied to the need for cognitive closure.
ReplyDeleteThis paper illuminated the intertwined relationship between categorization and cognition, suggesting they might be fundamentally the same process. This implies that without a thinking entity, categories can't exist, and cognition fundamentally arises from categorization. While on a theoretical level they seem to be identical, the challenge lies in understanding the formation and processing of categories and how cognitive functions shape and depend on categorization. What I'm trying to grasp is how categorization and cognition are identical. Is the assertion that they are the same in every respect, or is it suggesting that while categories form the basis of cognition, there exists some structural disparity?
ReplyDeleteI thought the fictional story (Section 17) regarding the man who got reverse amnesia and resultantly had infinite rote memory was a great analogy for what Professor Harnard has been talking about with ChatGPT and its limitations. While ChatGPT can help with summarizing basic information, it very much struggles with abstraction. This is because it possesses all of the knowledge of the “Big Gulp”, it stores information much like the man who cannot forget. However, it does not possess the ability to selectively detect invariance as it does not have sensorimotor grounding capabilities. This means that recognizing something and manipulating it (among the other things we can “do” with things) across different contexts is not possible, leading to the many limitations we get when attempting to use ChatGPT.
ReplyDeleteI was also at first confused as to how supervised learning would be more underdetermined than unsupervised learning - why can’t we just teach it how to categorize? But it is true that one thing could fit into any number of categorizations based on the selected features within different contexts. To label any individual categorization as right would be incredibly limiting, and with possibly infinite combinations simply training via error-corrective feedback is simply impossible.
This reading expanded my perspective when thinking about categorization. It nailed down the idea that it entails engaging in appropriate actions with the appropriate elements, and depending on what you, there will be specific consequences: positive and negative. This type of reinforcement allows our brain to be able to identify and extract the characteristics that differentiate categories. Building on the idea of categories, I do not 100% agree that categories are innate and that we are born with; however, I am sure that there are a few that are innate such as shapes. I think we are more so born with the proper tools to make categories and put things into them,and I believe it is more something that we learn (specifically the labels put on the categories) as we grow up and attend school. Additionally, I had a question/clarification about the meaning of vanishing intersections. From my understanding, it seems to pertain to the concept that certain categories are completely separate and do not share any common features, but I am unsure if that understanding is too concrete.
ReplyDeleteWhile reading this paper, I found myself particularly interested in sections 17 and 18. They made me think about people who have photographic memory. Whenever I have heard anything about photographic memory, I have only heard positive things about it, such as being able to perform very well in school due to remembering information in great detail. Would there also be some negative effects that come from this, similar to what we see in these sections, where issues arise with the inability to forget or ignore what makes every instant distinctive? How would this differ from the examples of Funes and “S”?
ReplyDeleteSection 7. on supervised learning struck me as a sort of optimization problem. Revolving around the optimization of one’s chicken sexing skills, making less miscategorizations as one moves up the ranks of white to yellow to orange…to brown and finally to black. I wonder how transfer learning comes into play here, may a black belt chicken sexer be better at categorizing male and female ducks?
ReplyDeleteMy stance is no different from the author that we could benefit most from an integration of cognitive neuroscience studies with evolution. In contrast to the previous paper on Evolution Psych, which emphasizes researching the causal factors, the highlight here is how neural mechanisms have changed over time, hence the much varied animal behaviours. Nonetheless, as mentioned in paper "Most cognitive tests are run under laboratory conditions to control confounding effects on cognition and yet the best estimates of fitness benefits should be measured in the wild where the importance of a specific cognitive ability will also depend on the environmental context." Balancing both sides in this field to delve deeper into cognition is certainly not a straightforward task.
ReplyDeleteFrom my perspective, our use of language involves patterns that are transmitted across time and space. Whenever we instruct or learn unconsciously, we create clusters using specific concepts, which enables us to amalgamate pre-existing feature detectors as the natural language's guide to combine different feature detectors.
ReplyDeleteNeural learning's word vector distance is a practical analogy of this: king + woman equals queen. Here, we have two grounded feature detectors, "king" and "woman," which can effectively classify this category. This classification results in a new classifier "queen," based on our use of natural language.
However, my concern arises if we consider scenarios where prior knowledge about the "king" and "woman" detectors was not attained. If that is the case, does the system fail to comprehend the "queen" category? Or in other word, not grounded if one don't have grounded cluster of king and woman？
This is my concern for such system where it fails to encapsulate the complexity of human's inner think network.
Human cognition goes beyond neural networks and words vectors. One of my major reservations lies in the fact that although language might provide cues, the feature detector present in the human brain doesn't grasp the concept of 'gradient.' A category might not merely be an assimilation or partition of several traits from other categories but more likely has a more intricate mechanism underlying it.
DeleteFor instance, the existence of categories that can be merged together holds relevance; however, each combination has a unique significance attached to it. The real intrigue lies in understanding how such amalgamations of existing categories into novel ones can occur. How can the processes in the brain that are responsible for these abstractions be quantified? Also, different feature detectors may not be eqully contribute when combine them to discover new case senarior.
I found this reading to be super interesting and insightful in its claim that cognition is categorization. The section on learned categories and supervised learning was of most interest to me as throughout my entire time in this course, I had struggled with the concepts related to learning. Specifically with regards to ChatGPT, my friends and I had repeatedly discussed after lectures about whether ChatGPT was actually indistinguishable from a human if we were to forget about the feeling aspect (HP). I had talked about how ChatGPT is able to learn also through a similar trial and error process in which users are able to teach ChatGPT and guide it through feedback to help answer further questions. Since ChatGPT does actually take into consideration user input to learn from its mistakes, it is very well capable of learning in this essence. From this reading and its usage of the grandmaster chicken sexers example, I can see that categorization is a mostly sensory skill that must be learned. I guess combining these two thoughts, my question is whether ChatGPT if given the sensorimotor capacities would be able to not only learn how to categorize but execute categorization to a level that we would not be able to distinguish from a human (T3).
ReplyDeleteThis comment has been removed by the author.
ReplyDeleteThis comment has been removed by the author.
ReplyDeleteI, along with many others, love the short story of Funes and have just read it this past week. The short story offers a great insight into the importance of categorization being an integral (if not THE integral) part of cognition and is a wonderful place to start exploring just how deep the rabbit hole goes. Funes or a Funes-like person could not exist as has been pointed out, especially not one who talks and has a grasp of language.
ReplyDeleteThe work of what goes on inside our heads is truly much more then just a copy-theory of cognition in which what’s in our heads is somehow a ‘representation’ of what’s out in the world, and Funes’ hypothetical existence shines a light on that. We necessarily prune our sensory experiences because if left unpruned we’d lose any cemblance of human thought; an uncategorized world where meaning cannot be abstracted and only pure sense experience is accessible and forever accessible.
Having read through all the replies I feel I don’t have much to add reguarding how we define categorization and the different fundemental ways to learn.
I very much appreciate the kid-sibly laying out of how to differentiate simly a dynamical system (i.e sand blowing in the wind) to an autonomous sensorimotor system. The lack of weaseling places a stable foundation on which the rest of the paper can build.
The fifth part about innate categories responds to my comment on the previous reading. An evolutionary set of prior categories must already exist in our minds when we are born, categories that are related to survival. These are all physical, however. The categorization of abstract things is very interesting to me. Abstraction is the path we take. As I thought more about abstract things such as “beauty” and “hope” and what they actually are, I grew more understanding of the subject. Beauty is simply a categorization of features that the person finds pleasing and hope is a response to future possibilities and wishes. Abstract things are reflections or compilations of multiple sensory categories and other abstract categories. At least that is how I think we categorize abstract categories from what I've understood from the reading.
ReplyDeleteMuch like the challenge of consciousness, categorization seems to remain elusive when it comes to introspection. While we can readily communicate the categories we create for specific inputs, the inner workings of our discrimination mechanism, the 'how' of it all, remains hidden. From what I understand, this process hinges on the absolute judgment of an object, relying on abstract features that can vary depending on the objects it's compared to, ultimately shaping the contextual significance of these features. In my view, when Harnad discusses the adaptive advantage of language addressing the limits of hearsay, it is pivotal in this context as it underscores the challenge of elucidating how we categorize and its connection to cognition. This is one of many points in this course that have more clearly shown me the true work that the field of cognitive science has cut out for it.
ReplyDeleteI found this reading to be the one that resonated with me the most so far. Categorization, what it is, how it occurs, and what it entails, is one of the concepts I've spent a lot of time thinking about. In this text, I was curious about how abstract concepts would be handled, and this was revealed towards the end of the text. It is stated that abstract concepts generally become available to us through hearsay, where we learn about their affordances. It is also explained that any abstract concept can be broken down into a set of concrete concepts when examined deeply enough. It is my personal belief that everything available to us can be abstracted since having a category is the capacity for abstraction, reducing something to an essential set of features is necessary.
ReplyDeleteWhile all of this is explained in the text, I find myself wondering about how we come to have categories of purely abstract things. As I mentioned previously, we get them through hearsay, but this only addresses how such categories are propagated. What about their inception? What must have happened for the very first person to think of the concept of justice as "justice"? How can we, without any hearsay or specific sensorimotor experiences, create new abstract concepts? My intuitive guess is that it has something to do with reordering and rearranging features of existing concepts into something that, to the minds that consider it, did not exist before.
The idea that there are some Stimuli which we need to respond the same way reminded me of how useful it is for a variety of different life forms to know what to do with certain kinds of things. Even bacteria need to be able to form this sort of classification to know how to interact with their environments in order to stay alive. It is interesting that this exists on different levels of life rather than just the complexity of the level we are interested in (humans). The discussion round how categorization is affected by the culture/environment one was brought up in/exists in reminds me of the efficient coding hypothesis (which normally is concerned w encoding vision) because the elements of the world that we have to deal with (most-lest) have a role in constructing the categories we form and how we parse out stimuli into these categories
ReplyDeleteI think the story of Funes and the memory-artist “S” show the somewhat opposite situation of a pure-thinking computer. Funes and S can experience the physical world but they cannot detect recurrence and abstract from their experience. A computer, on the other hand, is based primarily on categorizing abstract information, but it cannot experience the physical world. This makes me think of a potential problem of a T3 robot. If we can create a T3 robot that is able to sense the external world as we do, while the robot still have the large memory capacity, would it become a real-life embodiment of Funes? Since a robot has a larger memory capacity than humans and it cannot forget what is stored in its memory spontaneously. Would we have to design the T3 so that it deletes certain information in its memory on a regular basis? If so, how would this facilitate its performance in categorizing? In other words, we know that the ability to detect recurrence is important for abstraction, categorization, recognition, etc., which requires one to be able to forget, but what is the deeper process that links memory and categorization?
ReplyDeleteI found the short story Funes de memorious to give a very enlightening perspective on cognition as categorization: In the story, Ireneo Funes possesses an extraordinary memory, capable of recalling every minute detail of every experience he has ever had. However, Funes' inability to abstract or generalize from these details highlights a critical aspect of human cognition: the necessity of categorization. Funes' struggle symbolizes the limitations of a mind that can't categorize. Despite his incredible memory, he can't engage in higher-level thinking or abstract reasoning because he's unable to group and classify his endless stream of detailed observations. This inability to categorize renders his vast memory somewhat futile, as he can recall every detail but can't interpret or give them broader meaning.
ReplyDeleteFunes’ story started with his fall from a horse after which instead of impacting his memory capacity like most cases of head trauma, his memory was actually increased to the point that he remembered every detail that he encountered. Having followed the course, we know that in order to categorize, that is do the right thing with the right kind of thing which allows individuals to navigate the world, you need to be able to abstract the invariant features. These must be similar enough in two members that you can tell they are in the same category, but different enough that you can distinguish two objects as members of two different categories. In real life, Funes wouldn’t have been able to live in the way we do. Firstly, because he can’t do the right thing with the right kind of thing, so for example he can’t pour himself a cup of coffee, and secondly because he is not able to forget some rules in the case of language. In language, some rules need to be forgotten in specific cases, and in the case of hypermnesia forgetting is not possible so overgeneralization will most likely happen.
ReplyDelete