What exactly is suggested by Chomsky and "extreme nativists" when they claim certain capacities are "shaped neither by learning nor by evolution"?
If we accept that humans branched off from apes some eons ago, what could have happened in those thousands and thousands of years aside from evolution which would result in our current "innate" capacities? The only explanation seems to be some direct influence by an outside force about which we have no knowledge or even inkling of. Why do they deny these capacities from being products of evolution?
On the idea that categorization is abstraction, what is the process that occurs when comparing things based only on a single, abstract feature? For example, when given a perfect, blue circle and another shape, which is very nearly a perfect circle but would be categorized more as an ellipse, how does one determine which shape is more round? Perhaps some features, such as roundness, are of "higher levels" than other features. That is to say, roundness can be broken down into a number of other component features, such as the curvature of the lines and the smoothness of the surface. Does it entail that there are primitive features from which all other features are composed of?
While it may not answer your question exactly, I think you are leaning towards the ‘Geon Theory’ or Recognition by Components theory, proposed by Biederman. The paper is quite interesting (Biederman, I. (1987) Recognition-by-components: a theory of human image understanding. Psychol Rev. 1987 Apr;94(2):115-147) and he proposes that objects are made up of components, called geons, which is what allows us to visualize and identify scenes quickly. He doesn’t say there is only one primitive features from which everything else follows, but that there are fewer than 36 of these geons – cylindrical shapes- that can in different combinations make up the objects we see. I highly suggest you read it, you might enjoy it.
Categorization, as defined in class, is our ability to do the right thing with the right kind of thing. If I had to summarize the categorization process in a little more detail and in one sentence I would say that (1) our sensorimotor systems detect things in our environment, (2) a combination of unsupervised learning mechanisms (which cluster things based on augmented similarities and contrasts) and supervised learning mechanisms (which provide error-corrective feedback) extract the relevant features from these things, and (3) we selectively abstract these features into categories.
This sequence makes sense for the majority of our categories, which are learned. However, the ability to categorize linguistic input as compliant with Chomsky’s Universal Grammar (UG) was briefly discussed as one potentially innate category. Because children do not learn their native language on the basis of error-corrective feedback, the sequence above doesn’t hold for this kind of category formation. I don’t really understand how Chomsky can claim that UG wasn’t shaped by evolution at all, and claim that the rules are just a product of human brain structure/physiology, given that the human brain structure evolved from other primates’ (who seem perfectly able to categorize different communication signals not wholly unlike ours) brain structures.
Prof. Harnad's argument that categorization - doing the right stuff with the right kind of stuff - is what underlies cognition brought to mind the theories of the cognitive linguist George Lakoff. Lakoff writes on the roles of the metaphor and category formation as things that underlie and modify cognition.
Both Lakoff and Harnad seem to emphasize the importance of "grounding" - the interaction of the body with its environment - to cognition. In section 27 of the article, Harnad explores how language and communication gives humans a means of learning about a concept - a novel concept - by using knowledge and (this is essential) previously grounded (often lower-level) information. Lakoff in turn argues that concepts are created through a combination of linguistic experiences, historical information and interactions between the body and the environment.
I do think I'm struggling with how those grounded ideas might interact and intersect to form higher level or increasingly abstract concepts and categories. Are categories such those - especially ones taught largely through language - somewhat removed from sensorimotor inputs? If yes - how would increased sensorimotor interaction change the one's categorization of certain things?
We perceive things in our world by extracting the invariance of different phenomena, based on the features these phenomena afford to our sensory and motor systems. We do this via categorization: a dynamic cognitive system that takes a certain kind of input, and then performs a certain output. Based on supervised, unsupervised, and explicit learning, we continually update our rules for which inputs pertain to which outputs.
By studying our five sensory systems, we know how we detect absolute differences at the neurobiological level, but we don’t know how it is it that we detect kinds of inputs and how we respond differentially to them.
One thing that was clear from this paper, is that we lack a good definition of abstract, higher-order forms of categorical learning. The definition given for unsupervised learning is “sensory-motor mechanisms that cluster things according to their structural similarities, enhancing both the similarities and the contrasts.” But as is stated in section 16 and 27, this description doesn’t describe MOST of the categorical learning that we can do- it really only describes the level of learning toddlers are engaged in babbling around on the floor. By this definition, Monkeys are great at unsupervised learning. What is abstract learning then? Another definition given gets at it more fully, but more vaguely: “sensory motor mechanisms sample the structure and the correlations (including covariance and invariance under dynamic sensory motor transformations).” What constitutes a “dynamic sensory motor transformation”? I’m convinced by the Funes thought experiment that abstraction has to do with our ability to forget unimportant features. We are somehow able to abstract certain features as privileged, and then form a thought-constellation between those privileged features in our environment to features available in our memory.
While these are rational steps forward, we still don’t know what is it that separates our capabilities for categorization from the abilities of the apes. We don’t know what it means to form abstract categories on a neurological level. Clearly, our capacity for abstract categorization is coupled with our abilities for language… but how?
I went to a cognitive science talk at the Neuro last week, and a scientist named Katherine Duncan presented research on priming novelty vs. familiarity in memory formation. She found that when subjects were primed with novel stimuli, they were better able to discriminate between stimuli on a subsequent tasks. When subjects were primed with familiar stimuli, they were better able to group together stimuli and form associations on subsequent tasks. This research speaks to the importance of internal states in categorical learning and memory formations. Concerning abstract learning, I think it is important to consider what a stimuli affords our memory, given our mood state and our motivation in investigating the stimuli. Emotional/ motivational states could do large part in differentially weighing the detected features of stimuli. Our internal state could speak to the credit assignment problem in that it serves to provide the rules among many possibilities. However, I admit I have no idea what I am talking about what I say “internal state”. Obviously we have a lot of work to do in this arena if it is to impact the credit assignment problem and inform our understanding of how it is that we categorize.
I was a bit misleading in my description above. Both supervised and unsupervised learning can be performed by apes, and both are done via corrective feedback. Unsupervised learning is completed via sensory-motor feedback (nourishment after eating good mushroom, stomach pain after eating a bad one) and supervised is when someone else provides the feedback to our behaviour. The different between ourselves and apes is the capacity for explicit learning, as well as unsupervised and supervised learning at a more abstract level. Both humans and apes surely have neural nets that are able to render some features of some categories more salient, more prominent, based on the internal state and our surrounding environment, but somehow, via language, we are able to do this with many layers of abstraction and apes are not. We are able to categorize, and categorically perceive at levels greatly abstracted from the original sensory-motor experiences we had with any given thing. To show how complicated this is, Wittgenstein's "word meaning is word use" shows us that the meaning of words that are more removed from sensory-motor experience are largely tied to their use in culture. The word "Trump" largely primes different features of different categories for Texans than it does for San Franciscans.
Regarding: “And just as there is compression within each color range, there is expansion between them: Equal-sized frequency differences look much smaller and are harder to detect when they are within one color category than when they cross the boundary from one category to the other.”
This made me remember a simple colour perception task I read about in one of my previous courses, which looked at English vs. Russian speakers to investigate whether our perceptual processes are influenced by how we linguistically carve up the colour spectrum (Winawer et al. 2007). During this task, the subjects were presented with a blue square, which disappeared and was followed by two more blue squares after a brief pause. The subjects then had to determine which of the two squares matched the colour of the original stimulus. Importantly, the Russian language makes a linguistic distinction between lighter shades of blue “goluboy” and darker shades of blue “siniy”. Winawer et al. found that the Russian-speaking subjects had significantly faster reaction times when the following two stimuli they had to choose from were of different categories (i.e. goluboy-siniy), compared to when they where of the same category. English-speaking participants showed no difference in reaction time between these conditions, suggesting that the linguistic terms available to speakers of a given language can facilitate categorical perception (reminiscent of the Whorf hypothesis).
For Russian speakers, could this be thought of in terms of a recoding process whereby within the already compressed category of “blue”, light and dark are further compressed (individually) and separated by a boundary region similar to—but less pronounced than—the boundary regions which separate regular colour categories?
Winawer, Jonathan, et al. "Russian blues reveal effects of language on color discrimination." Proceedings of the National Academy of Sciences 104.19 (2007): 7780-7785.
I'm a native Russian speaker so I was intrigued by this comment. I think a reason why it takes longer to make a linguistic distinction between two stimuli from the same category of blue could be due to perception and each stimulus' relation to the other. For example, "goluboy" is generally used for sky blue and lighter shades of blue, while "siniy" is generally used for your standard blue and anything darker, such as navy. Both terms still encompass a wide range of hues and shades of blue. As such, if I were asked to distinguish between blue and navy, I would probably have to think about whether to still call both of them "siniy", or call the blue one lighter blue ("cvetlo ciniy) and the navy one darker blue ("tomno siniy"). Another (less likely) option in this scenario is whether, compared to the navy, the regular blue can be considered "goluboy" instead.
In short, I think that, for myself and probably other (but maybe not all) Russian speakers, when relating shades of blue to one another, the already compressed categories of light and dark blue can also be separated by light and dark shades of light blue, and light and dark shades of dark blue, respectively.
RE: “…complex configurations of sand that merely respond (and have always responded) differentially to different kinds of input in the way ordinary sand responds (and has always responded) to wind from different directions -- is to show that at one time it was not so: that it did not always respond differentially as it does now.”
I’m not sure I understand how this sand example exemplifies that “categorization is intimately tied to learning.” I understand that the example is demonstrating how sand responds differently to different kinds of input which shows that it is adaptive however, is it possible for an inanimate object to “learn?” What is the definition of learning in this sense? Would it be the same to say that categorization is tied to adaptation, instead of learning? I guess I just need more clarification as to what it means for categorization to be tied to learning.
It is much easier for me to understand what learning is after reading the part on supervised learning. It is interesting to think of categorization as a skill as opposed to something that you are born with as it seems as though the majority of the population are able to do so. However, I wonder in cases where someone is in a tragic accident that damages their brain, some of these patients have difficulty categorizing objects. Would this suggest that categorization is actually developed in the brain? Or perhaps it is because the patient has already learned the skill of categorization and the brain has designated a part of it accordingly.
Hi Jenna, there are actually brain areas associated for specific types of categorization such as face perception and people with damage to that area do show prosopagnosia. Even for categories that are not so well localized and specific, I think it can be argued that our brains are still biased towards learning categories and so that we can establish boundaries and associatively learn to group things together (our fallible memories that allow us to forget details would be an example).
Harnad describes the ways in which we categorize inputs in our environment and the importance this has for all kinds of human behaviours and reactions to our surroundings and indeed its absolute importance to cognition as a whole. He defines categorization as “any systematic differential interaction between an autonomous, adaptive sensorimotor system and its world.”
I find it interesting to think about why categorization - an ability that develops from infancy really - can be so so difficult to replicate in machines. I think the reason it is so difficult to replicate is because it depends on the ability to tune out a large amount of input so as to focus on the key attributes of an object (ie the “affordances”) which I imagine machines are not that good at doing. This is essential in order to properly recognize objects as belonging to a certain category, even when there is variation between those objects. One can imagine two chairs that differ in any number of ways (size, colour, location, material, etc) but we correctly identify them as “chairs” because they “afford” us the ability to sit. As well I think the word “adaptive” is a key aspect of Harnad’s definition and highlights how humans, through corrective feedback, can learn to adjust their categorical perceptions by assigning weight to different features. In particular, I think it is interesting to contrast examples in which this categorical perception is likely innate (e.g. colour perception, universal grammar) with categories that are learned through experience. To explore which types of categorical perceptions are innate or else can be learned without corrective feedback would provide interesting insight into human development and cognition.
This article also makes me think about the complexity of visual processing (ie in perception classes we learn that the eye does not “see”, it is the brain that “sees”) and how this sensory processing is intimately linked with other cognitive capacities allowing us to seamlessly interact with our environment. Fune’s story shows us how categorization is inherently adaptive and systematic and necessary in order to effectively interact with our environment. In comparison to vision, a camera represents images as a series of coloured pixels. A camera cannot “see” (unless it has software to help it do so). The process of seeing requires abstraction from the sensorimotor input (ie the pixels of lights) in order to see or “recognize” shapes and objects. From there it is possible to recognize those objects as belonging to different categories (for example “rose”). If the flower was a rose you might also abstract out other qualities such as “red” or the broader category of “flowers” or “living things”. I think this highlights different levels of abstraction from identical sensorimotor inputs. One could propose that abstract concepts (with less objective material basis) such as emotions or concepts of love, truth, beauty etc are just a higher level category of abstraction from lower levels of abstractions originating in sensorimotor inputs.
Hi Chloe, I think you are right in suggesting that tuning out large quantities of inputs helps the process of categorization. I also think that pattern recognition, which can be developed in AI, plays a vital role in categorization. I friend of mine who works at google told me that they are not trying to make machines cognize but are trying to make machines recognize as many patterns as possible. For professor Harnad learning is based on adaptive systems with real time history. However, pattern recognition could possibly replace the bottom up affordance based categorization process. An example would be Facebooks ability to detect suicideality based on photographs and other facebook posts. This is a complex task which requires a lot of training to detect. The fact that a computer program can detect something like this is fascinating and calls into question the bottom up categorization process.
Hi Soham, I agree that pattern recognition would be a key element in developing a capacity for categorization in AI. Face recognition software used on multiple social media platforms (Snapchat filters are a good example)also came to mind and how this has advanced impressively mimicking some aspects of human visual perception. Detecting a complex behaviour such as suicidality is particularly interesting - I wonder what key words or images they use to detect this? However I am also wondering why you think this calls into question bottom up categorization processes? I think they both depend on a process of "pattern recognition" which is essentially the ability to use specific features to distinguish between members of the category and non-members (in your example it would be Facebook activity indicating suicidality and all other Facebook activity). My point about affordances is that humans will often focus on (and abstract) the features of an object that "afford" us something in order to make categorizations. For example a baby would be quick to distinguish between objects that give food and those that do not. I think it would be more difficult for a computer to pick out these affordances simply because they don't have these innate needs. As a result much of the learning that is intuitive for humans (and many animals) would need to be explicitly taught to computers in the form of algorithms. (Although I do agree this is by no means an insurmountable obstacle!)
On (27) The Adaptive Advantage of Language: Hearsay
This is an interesting application of supervised learning, in that the explicit instructions given by a master chicken-sorter are feedback on the unfolding CP learning happening in the chicken-sorting learners. Language, though it cannot perfectly capture our implicit categorizations (as evident in the fact that people misreport or underreport the invariants they use for categorizing), is an effective and efficient tool for learning and teaching, especially when it comes to providing corrective feedback.
I wonder if other people with synesthesia (like the stage memory-artist “S”) also have better memory than the norm? What about synesthesia would enhance your memory? Just because everything is richer in experience? Could you not say then that any sort of multimodal experience would be more memorable than unimodal experiences? Also regarding S, I find it interesting that he has trouble with abstract concepts. Is it because you can’t visualize abstract concepts in a physical manner? I wonder if they would be able to remember dreams equally as well as actual experiences they had. I also wonder how these people are when it comes to emotions.
RE: Nativism about categories and Universal Grammar (UG)
Although nativism about categories is an extreme stance, I believe an alternate take that preserves the notion that our brains are “pre-structured,” but also recognizes individual differences in perception and learned categories, could be believable. The alternate supports that the brain is pre-restructured, and that we are born with innate abilities to detect many types of invariances (more than we can now) based on their physical properties and on the limitations of our biology. As the brain develops and learns, undergoing various linguistic/ sensorimotor experiences, our ability to detect particular types of invariances are made stronger and categories are reinforced, while others don’t get expressed potentially crippling that capacity. So, specific categorization capacities aren’t prewired, but rather the capacities that afford our ability to categorize stimuli with certain properties is. Whatever specific function or category ends up taking advantage of the ability is a result of evolution and learning. So, when Chomsky says we have an innate capacity specific for language, I would agree that the capacity is innate, but disagree that it’s specifically “for language.” There is research that suggests language stems out of other abilities. Such as research which demonstrates that the brain structures implied in UG are also implied in categorization associated with music perception, like pitch, prosody and beat recognition, and postulate that musical capacities evolutionarily kicked off language, and that music perception underlie, support and precede linguistic perception. So, I don’t think we have any innate categorization capabilities for X, where X is a specific function that relates to some brain region, but that the structure of our brain gives us innate capacities to categorize and work with certain properties of stimuli that lead to categories and categorization abilities. In regards to the role of language in categories and CP, by this perspective language can influence categorization and categories, but in a way that our innate abilities allow for. So, language will never get us to perceive a new color, Blorf, but may allow us to create a new animal Blorf.
I feel like the capacity to categorize is innate, but the rest is learned. We began by categorizing things that were evolutionarily important. It was important for our survival. Then, as language evolved and we had more complex objects in our lives that had many examples of the same kind, and we had to name and categorize them. We could not have foreseen that we would need categorizing for different kinds of furniture, but we always knew that we would need categorization. Thus, is it not the ability to categorize that is innate, and the categories are then learned as needed? Only a few exceptions would exist such as color. Since we are very plastic in our ability to change and adapt, I believe it would not be hard for the brain to learn to categorize things that began to be more complex and abstract with language. Although they are closely linked, I think they are different. I don’t see how Fodor can make the point of categories being innate.
I am also very confused about Funes the memorious. Even if he remembers everything, like people with photographic memories, why can he not abstract? Surely he can recognize without forgetting about all the other possibilities. He just would need to ignore it for a second, which he must be doing anyway because all that information in his brain cannot be all salient at the same time. I feel like forgetting is a very different ability from ignoring. Did he not learn that some features were to be given more weight than the others? He must have, so he must be able to abstract. On the other hand, some categories we learn must confuse him more as they add to the swan and the duckling having more in common. However, if you were to show the animals to Funes as a baby, he would have arbitrarily selected the most salient visual cues and abstracted.
Deniz, while there are many things that we certainly have evolutionarily innate capacities to detect with obvious evolutionary narratives behind them (faces, snakes, spiders, good/bad smells, etc), I think it is important to remember that evolution also affords us domain-general categorization capacities since evolution cannot really know what kind of environment we will find ourselves in. There is even evidence of this in bees who recognize bee-keeper faces, despite bee-keepers not really being a part of the bee's evolutionary development. Therefore I think that we possessed domain-general categorization processes long before we ever evolved language to add to our acultured categorization processes.
I found the Funes and S examples interesting when discussed in the context of language acquisition. For these two individuals, every input is unique and memorized as such. From my understanding, when children acquire language they initially memorize their input just like Funes and S. As their exposure increases, they begin to make abstractions about the patterns in their input to acquire rules that they can apply to novel input. This concept can be seen in the U-shaped development of irregular verbs. Children first hear and memorize the conjugated verbs “broke” (irregular) and “played” (regular). As their exposure to conjugated regular verbs increases, they acquire the rule of adding –ed for past tensed verbs. Although children would never hear the word “breaked” as input, their production of this conjugation suggests they have acquired the rule for past tense. Lastly, children will recognize that there are exceptions to the rule that must be learned one at a time and will return to their initial production of “broke”. In this example, categorization, as defined by our ability to do the right thing with the right kind of thing, depends on a child’s ability to ignore the differences between words allow them to form rules about word forms. However, it also depends on recognizing key differences to learn the exceptions of input that need to be memorized as unique.
“According to Chomsky, our capacity to detect and generate UG-compliant strings of words is shaped neither by learning nor by evolution; it is instead somehow inherent in the structure of our brains as a matter of structural inevitability.”
I believe this is a mistaken reading of Chomsky. Perhaps this is true of early Chomsky, but from the 1980s onward Chomsky and his followers have been largely concern with the principles and parameters approach to UG and the minimalist program. These two elements are part of the biolinguistic program that seeks to find a biological and evolutionary explanation for how UG could have arisen (i.e. Chomsky, Hauser Fitch 2002). Moreover, the principles and parameters framework (Chomsky 1981, Chomsky and Lasnik 1993) posits that Universal Grammar includes a finite number of parameter/switches (i.e. choosing between SVO and SOV languages, or head-initial or head-final languages), that the language learning infant learn to code switch early on to rule out possible grammar that are not relevant to the language(s) they are acquiring. Therefore within Chomsky’s theories, at least for the last 37 years, our capacity to generate UG-compliant strings is shaped BOTH by learning and by evolution.
What exactly is suggested by Chomsky and "extreme nativists" when they claim certain capacities are "shaped neither by learning nor by evolution"?
ReplyDeleteIf we accept that humans branched off from apes some eons ago, what could have happened in those thousands and thousands of years aside from evolution which would result in our current "innate" capacities? The only explanation seems to be some direct influence by an outside force about which we have no knowledge or even inkling of. Why do they deny these capacities from being products of evolution?
On the idea that categorization is abstraction, what is the process that occurs when comparing things based only on a single, abstract feature? For example, when given a perfect, blue circle and another shape, which is very nearly a perfect circle but would be categorized more as an ellipse, how does one determine which shape is more round? Perhaps some features, such as roundness, are of "higher levels" than other features. That is to say, roundness can be broken down into a number of other component features, such as the curvature of the lines and the smoothness of the surface. Does it entail that there are primitive features from which all other features are composed of?
ReplyDeleteWhile it may not answer your question exactly, I think you are leaning towards the ‘Geon Theory’ or Recognition by Components theory, proposed by Biederman. The paper is quite interesting (Biederman, I. (1987) Recognition-by-components: a theory of human image understanding. Psychol Rev. 1987 Apr;94(2):115-147) and he proposes that objects are made up of components, called geons, which is what allows us to visualize and identify scenes quickly. He doesn’t say there is only one primitive features from which everything else follows, but that there are fewer than 36 of these geons – cylindrical shapes- that can in different combinations make up the objects we see. I highly suggest you read it, you might enjoy it.
DeleteCategorization, as defined in class, is our ability to do the right thing with the right kind of thing. If I had to summarize the categorization process in a little more detail and in one sentence I would say that (1) our sensorimotor systems detect things in our environment, (2) a combination of unsupervised learning mechanisms (which cluster things based on augmented similarities and contrasts) and supervised learning mechanisms (which provide error-corrective feedback) extract the relevant features from these things, and (3) we selectively abstract these features into categories.
ReplyDeleteThis sequence makes sense for the majority of our categories, which are learned. However, the ability to categorize linguistic input as compliant with Chomsky’s Universal Grammar (UG) was briefly discussed as one potentially innate category. Because children do not learn their native language on the basis of error-corrective feedback, the sequence above doesn’t hold for this kind of category formation. I don’t really understand how Chomsky can claim that UG wasn’t shaped by evolution at all, and claim that the rules are just a product of human brain structure/physiology, given that the human brain structure evolved from other primates’ (who seem perfectly able to categorize different communication signals not wholly unlike ours) brain structures.
Prof. Harnad's argument that categorization - doing the right stuff with the right kind of stuff - is what underlies cognition brought to mind the theories of the cognitive linguist George Lakoff. Lakoff writes on the roles of the metaphor and category formation as things that underlie and modify cognition.
ReplyDeleteBoth Lakoff and Harnad seem to emphasize the importance of "grounding" - the interaction of the body with its environment - to cognition. In section 27 of the article, Harnad explores how language and communication gives humans a means of learning about a concept - a novel concept - by using knowledge and (this is essential) previously grounded (often lower-level) information. Lakoff in turn argues that concepts are created through a combination of linguistic experiences, historical information and interactions between the body and the environment.
I do think I'm struggling with how those grounded ideas might interact and intersect to form higher level or increasingly abstract concepts and categories. Are categories such those - especially ones taught largely through language - somewhat removed from sensorimotor inputs? If yes - how would increased sensorimotor interaction change the one's categorization of certain things?
This comment has been removed by the author.
ReplyDeleteThis comment has been removed by the author.
ReplyDeleteWe perceive things in our world by extracting the invariance of different phenomena, based on the features these phenomena afford to our sensory and motor systems. We do this via categorization: a dynamic cognitive system that takes a certain kind of input, and then performs a certain output. Based on supervised, unsupervised, and explicit learning, we continually update our rules for which inputs pertain to which outputs.
ReplyDeleteBy studying our five sensory systems, we know how we detect absolute differences at the neurobiological level, but we don’t know how it is it that we detect kinds of inputs and how we respond differentially to them.
One thing that was clear from this paper, is that we lack a good definition of abstract, higher-order forms of categorical learning. The definition given for unsupervised learning is “sensory-motor mechanisms that cluster things according to their structural similarities, enhancing both the similarities and the contrasts.” But as is stated in section 16 and 27, this description doesn’t describe MOST of the categorical learning that we can do- it really only describes the level of learning toddlers are engaged in babbling around on the floor. By this definition, Monkeys are great at unsupervised learning. What is abstract learning then? Another definition given gets at it more fully, but more vaguely: “sensory motor mechanisms sample the structure and the correlations (including covariance and invariance under dynamic sensory motor transformations).” What constitutes a “dynamic sensory motor transformation”? I’m convinced by the Funes thought experiment that abstraction has to do with our ability to forget unimportant features. We are somehow able to abstract certain features as privileged, and then form a thought-constellation between those privileged features in our environment to features available in our memory.
While these are rational steps forward, we still don’t know what is it that separates our capabilities for categorization from the abilities of the apes. We don’t know what it means to form abstract categories on a neurological level. Clearly, our capacity for abstract categorization is coupled with our abilities for language… but how?
I went to a cognitive science talk at the Neuro last week, and a scientist named Katherine Duncan presented research on priming novelty vs. familiarity in memory formation. She found that when subjects were primed with novel stimuli, they were better able to discriminate between stimuli on a subsequent tasks. When subjects were primed with familiar stimuli, they were better able to group together stimuli and form associations on subsequent tasks. This research speaks to the importance of internal states in categorical learning and memory formations.
Concerning abstract learning, I think it is important to consider what a stimuli affords our memory, given our mood state and our motivation in investigating the stimuli. Emotional/ motivational states could do large part in differentially weighing the detected features of stimuli. Our internal state could speak to the credit assignment problem in that it serves to provide the rules among many possibilities. However, I admit I have no idea what I am talking about what I say “internal state”. Obviously we have a lot of work to do in this arena if it is to impact the credit assignment problem and inform our understanding of how it is that we categorize.
I was a bit misleading in my description above. Both supervised and unsupervised learning can be performed by apes, and both are done via corrective feedback. Unsupervised learning is completed via sensory-motor feedback (nourishment after eating good mushroom, stomach pain after eating a bad one) and supervised is when someone else provides the feedback to our behaviour. The different between ourselves and apes is the capacity for explicit learning, as well as unsupervised and supervised learning at a more abstract level. Both humans and apes surely have neural nets that are able to render some features of some categories more salient, more prominent, based on the internal state and our surrounding environment, but somehow, via language, we are able to do this with many layers of abstraction and apes are not. We are able to categorize, and categorically perceive at levels greatly abstracted from the original sensory-motor experiences we had with any given thing.
DeleteTo show how complicated this is, Wittgenstein's "word meaning is word use" shows us that the meaning of words that are more removed from sensory-motor experience are largely tied to their use in culture. The word "Trump" largely primes different features of different categories for Texans than it does for San Franciscans.
Regarding: “And just as there is compression within each color range, there is expansion between them: Equal-sized frequency differences look much smaller and are harder to detect when they are within one color category than when they cross the boundary from one category to the other.”
ReplyDeleteThis made me remember a simple colour perception task I read about in one of my previous courses, which looked at English vs. Russian speakers to investigate whether our perceptual processes are influenced by how we linguistically carve up the colour spectrum (Winawer et al. 2007). During this task, the subjects were presented with a blue square, which disappeared and was followed by two more blue squares after a brief pause. The subjects then had to determine which of the two squares matched the colour of the original stimulus. Importantly, the Russian language makes a linguistic distinction between lighter shades of blue “goluboy” and darker shades of blue “siniy”. Winawer et al. found that the Russian-speaking subjects had significantly faster reaction times when the following two stimuli they had to choose from were of different categories (i.e. goluboy-siniy), compared to when they where of the same category. English-speaking participants showed no difference in reaction time between these conditions, suggesting that the linguistic terms available to speakers of a given language can facilitate categorical perception (reminiscent of the Whorf hypothesis).
For Russian speakers, could this be thought of in terms of a recoding process whereby within the already compressed category of “blue”, light and dark are further compressed (individually) and separated by a boundary region similar to—but less pronounced than—the boundary regions which separate regular colour categories?
Winawer, Jonathan, et al. "Russian blues reveal effects of language on color discrimination." Proceedings of the National Academy of Sciences 104.19 (2007): 7780-7785.
Hi Teresa,
DeleteI'm a native Russian speaker so I was intrigued by this comment. I think a reason why it takes longer to make a linguistic distinction between two stimuli from the same category of blue could be due to perception and each stimulus' relation to the other. For example, "goluboy" is generally used for sky blue and lighter shades of blue, while "siniy" is generally used for your standard blue and anything darker, such as navy. Both terms still encompass a wide range of hues and shades of blue. As such, if I were asked to distinguish between blue and navy, I would probably have to think about whether to still call both of them "siniy", or call the blue one lighter blue ("cvetlo ciniy) and the navy one darker blue ("tomno siniy"). Another (less likely) option in this scenario is whether, compared to the navy, the regular blue can be considered "goluboy" instead.
In short, I think that, for myself and probably other (but maybe not all) Russian speakers, when relating shades of blue to one another, the already compressed categories of light and dark blue can also be separated by light and dark shades of light blue, and light and dark shades of dark blue, respectively.
RE: “…complex configurations of sand that merely respond (and have always responded) differentially to different kinds of input in the way ordinary sand responds (and has always responded) to wind from different directions -- is to show that at one time it was not so: that it did not always respond differentially as it does now.”
ReplyDeleteI’m not sure I understand how this sand example exemplifies that “categorization is intimately tied to learning.” I understand that the example is demonstrating how sand responds differently to different kinds of input which shows that it is adaptive however, is it possible for an inanimate object to “learn?” What is the definition of learning in this sense? Would it be the same to say that categorization is tied to adaptation, instead of learning? I guess I just need more clarification as to what it means for categorization to be tied to learning.
RE: Supervised Learning
ReplyDeleteIt is much easier for me to understand what learning is after reading the part on supervised learning. It is interesting to think of categorization as a skill as opposed to something that you are born with as it seems as though the majority of the population are able to do so. However, I wonder in cases where someone is in a tragic accident that damages their brain, some of these patients have difficulty categorizing objects. Would this suggest that categorization is actually developed in the brain? Or perhaps it is because the patient has already learned the skill of categorization and the brain has designated a part of it accordingly.
Hi Jenna, there are actually brain areas associated for specific types of categorization such as face perception and people with damage to that area do show prosopagnosia. Even for categories that are not so well localized and specific, I think it can be argued that our brains are still biased towards learning categories and so that we can establish boundaries and associatively learn to group things together (our fallible memories that allow us to forget details would be an example).
DeleteHarnad describes the ways in which we categorize inputs in our environment and the importance this has for all kinds of human behaviours and reactions to our surroundings and indeed its absolute importance to cognition as a whole. He defines categorization as “any systematic differential interaction between an autonomous, adaptive sensorimotor system and its world.”
ReplyDeleteI find it interesting to think about why categorization - an ability that develops from infancy really - can be so so difficult to replicate in machines. I think the reason it is so difficult to replicate is because it depends on the ability to tune out a large amount of input so as to focus on the key attributes of an object (ie the “affordances”) which I imagine machines are not that good at doing. This is essential in order to properly recognize objects as belonging to a certain category, even when there is variation between those objects. One can imagine two chairs that differ in any number of ways (size, colour, location, material, etc) but we correctly identify them as “chairs” because they “afford” us the ability to sit. As well I think the word “adaptive” is a key aspect of Harnad’s definition and highlights how humans, through corrective feedback, can learn to adjust their categorical perceptions by assigning weight to different features. In particular, I think it is interesting to contrast examples in which this categorical perception is likely innate (e.g. colour perception, universal grammar) with categories that are learned through experience. To explore which types of categorical perceptions are innate or else can be learned without corrective feedback would provide interesting insight into human development and cognition.
This article also makes me think about the complexity of visual processing (ie in perception classes we learn that the eye does not “see”, it is the brain that “sees”) and how this sensory processing is intimately linked with other cognitive capacities allowing us to seamlessly interact with our environment. Fune’s story shows us how categorization is inherently adaptive and systematic and necessary in order to effectively interact with our environment. In comparison to vision, a camera represents images as a series of coloured pixels. A camera cannot “see” (unless it has software to help it do so). The process of seeing requires abstraction from the sensorimotor input (ie the pixels of lights) in order to see or “recognize” shapes and objects. From there it is possible to recognize those objects as belonging to different categories (for example “rose”). If the flower was a rose you might also abstract out other qualities such as “red” or the broader category of “flowers” or “living things”. I think this highlights different levels of abstraction from identical sensorimotor inputs. One could propose that abstract concepts (with less objective material basis) such as emotions or concepts of love, truth, beauty etc are just a higher level category of abstraction from lower levels of abstractions originating in sensorimotor inputs.
Hi Chloe,
DeleteI think you are right in suggesting that tuning out large quantities of inputs helps the process of categorization. I also think that pattern recognition, which can be developed in AI, plays a vital role in categorization. I friend of mine who works at google told me that they are not trying to make machines cognize but are trying to make machines recognize as many patterns as possible. For professor Harnad learning is based on adaptive systems with real time history. However, pattern recognition could possibly replace the bottom up affordance based categorization process. An example would be Facebooks ability to detect suicideality based on photographs and other facebook posts. This is a complex task which requires a lot of training to detect. The fact that a computer program can detect something like this is fascinating and calls into question the bottom up categorization process.
Hi Soham, I agree that pattern recognition would be a key element in developing a capacity for categorization in AI. Face recognition software used on multiple social media platforms (Snapchat filters are a good example)also came to mind and how this has advanced impressively mimicking some aspects of human visual perception. Detecting a complex behaviour such as suicidality is particularly interesting - I wonder what key words or images they use to detect this? However I am also wondering why you think this calls into question bottom up categorization processes? I think they both depend on a process of "pattern recognition" which is essentially the ability to use specific features to distinguish between members of the category and non-members (in your example it would be Facebook activity indicating suicidality and all other Facebook activity). My point about affordances is that humans will often focus on (and abstract) the features of an object that "afford" us something in order to make categorizations. For example a baby would be quick to distinguish between objects that give food and those that do not. I think it would be more difficult for a computer to pick out these affordances simply because they don't have these innate needs. As a result much of the learning that is intuitive for humans (and many animals) would need to be explicitly taught to computers in the form of algorithms. (Although I do agree this is by no means an insurmountable obstacle!)
DeleteOn (27) The Adaptive Advantage of Language: Hearsay
ReplyDeleteThis is an interesting application of supervised learning, in that the explicit instructions given by a master chicken-sorter are feedback on the unfolding CP learning happening in the chicken-sorting learners. Language, though it cannot perfectly capture our implicit categorizations (as evident in the fact that people misreport or underreport the invariants they use for categorizing), is an effective and efficient tool for learning and teaching, especially when it comes to providing corrective feedback.
I wonder if other people with synesthesia (like the stage memory-artist “S”) also have better memory than the norm? What about synesthesia would enhance your memory? Just because everything is richer in experience? Could you not say then that any sort of multimodal experience would be more memorable than unimodal experiences? Also regarding S, I find it interesting that he has trouble with abstract concepts. Is it because you can’t visualize abstract concepts in a physical manner? I wonder if they would be able to remember dreams equally as well as actual experiences they had. I also wonder how these people are when it comes to emotions.
ReplyDeleteRE: Nativism about categories and Universal Grammar (UG)
ReplyDeleteAlthough nativism about categories is an extreme stance, I believe an alternate take that preserves the notion that our brains are “pre-structured,” but also recognizes individual differences in perception and learned categories, could be believable. The alternate supports that the brain is pre-restructured, and that we are born with innate abilities to detect many types of invariances (more than we can now) based on their physical properties and on the limitations of our biology. As the brain develops and learns, undergoing various linguistic/ sensorimotor experiences, our ability to detect particular types of invariances are made stronger and categories are reinforced, while others don’t get expressed potentially crippling that capacity. So, specific categorization capacities aren’t prewired, but rather the capacities that afford our ability to categorize stimuli with certain properties is. Whatever specific function or category ends up taking advantage of the ability is a result of evolution and learning. So, when Chomsky says we have an innate capacity specific for language, I would agree that the capacity is innate, but disagree that it’s specifically “for language.” There is research that suggests language stems out of other abilities. Such as research which demonstrates that the brain structures implied in UG are also implied in categorization associated with music perception, like pitch, prosody and beat recognition, and postulate that musical capacities evolutionarily kicked off language, and that music perception underlie, support and precede linguistic perception. So, I don’t think we have any innate categorization capabilities for X, where X is a specific function that relates to some brain region, but that the structure of our brain gives us innate capacities to categorize and work with certain properties of stimuli that lead to categories and categorization abilities. In regards to the role of language in categories and CP, by this perspective language can influence categorization and categories, but in a way that our innate abilities allow for. So, language will never get us to perceive a new color, Blorf, but may allow us to create a new animal Blorf.
I feel like the capacity to categorize is innate, but the rest is learned. We began by categorizing things that were evolutionarily important. It was important for our survival. Then, as language evolved and we had more complex objects in our lives that had many examples of the same kind, and we had to name and categorize them. We could not have foreseen that we would need categorizing for different kinds of furniture, but we always knew that we would need categorization. Thus, is it not the ability to categorize that is innate, and the categories are then learned as needed? Only a few exceptions would exist such as color. Since we are very plastic in our ability to change and adapt, I believe it would not be hard for the brain to learn to categorize things that began to be more complex and abstract with language. Although they are closely linked, I think they are different. I don’t see how Fodor can make the point of categories being innate.
ReplyDeleteI am also very confused about Funes the memorious. Even if he remembers everything, like people with photographic memories, why can he not abstract? Surely he can recognize without forgetting about all the other possibilities. He just would need to ignore it for a second, which he must be doing anyway because all that information in his brain cannot be all salient at the same time. I feel like forgetting is a very different ability from ignoring. Did he not learn that some features were to be given more weight than the others? He must have, so he must be able to abstract. On the other hand, some categories we learn must confuse him more as they add to the swan and the duckling having more in common. However, if you were to show the animals to Funes as a baby, he would have arbitrarily selected the most salient visual cues and abstracted.
Deniz, while there are many things that we certainly have evolutionarily innate capacities to detect with obvious evolutionary narratives behind them (faces, snakes, spiders, good/bad smells, etc), I think it is important to remember that evolution also affords us domain-general categorization capacities since evolution cannot really know what kind of environment we will find ourselves in. There is even evidence of this in bees who recognize bee-keeper faces, despite bee-keepers not really being a part of the bee's evolutionary development. Therefore I think that we possessed domain-general categorization processes long before we ever evolved language to add to our acultured categorization processes.
DeleteI found the Funes and S examples interesting when discussed in the context of language acquisition. For these two individuals, every input is unique and memorized as such. From my understanding, when children acquire language they initially memorize their input just like Funes and S. As their exposure increases, they begin to make abstractions about the patterns in their input to acquire rules that they can apply to novel input.
ReplyDeleteThis concept can be seen in the U-shaped development of irregular verbs. Children first hear and memorize the conjugated verbs “broke” (irregular) and “played” (regular). As their exposure to conjugated regular verbs increases, they acquire the rule of adding –ed for past tensed verbs. Although children would never hear the word “breaked” as input, their production of this conjugation suggests they have acquired the rule for past tense. Lastly, children will recognize that there are exceptions to the rule that must be learned one at a time and will return to their initial production of “broke”.
In this example, categorization, as defined by our ability to do the right thing with the right kind of thing, depends on a child’s ability to ignore the differences between words allow them to form rules about word forms. However, it also depends on recognizing key differences to learn the exceptions of input that need to be memorized as unique.
“According to Chomsky, our capacity to detect and generate UG-compliant strings of words is shaped neither by learning nor by evolution; it is instead somehow inherent in the structure of our brains as a matter of structural inevitability.”
ReplyDeleteI believe this is a mistaken reading of Chomsky. Perhaps this is true of early Chomsky, but from the 1980s onward Chomsky and his followers have been largely concern with the principles and parameters approach to UG and the minimalist program. These two elements are part of the biolinguistic program that seeks to find a biological and evolutionary explanation for how UG could have arisen (i.e. Chomsky, Hauser Fitch 2002). Moreover, the principles and parameters framework (Chomsky 1981, Chomsky and Lasnik 1993) posits that Universal Grammar includes a finite number of parameter/switches (i.e. choosing between SVO and SOV languages, or head-initial or head-final languages), that the language learning infant learn to code switch early on to rule out possible grammar that are not relevant to the language(s) they are acquiring. Therefore within Chomsky’s theories, at least for the last 37 years, our capacity to generate UG-compliant strings is shaped BOTH by learning and by evolution.