|
|
Post by kjholste on Mar 27, 2016 4:09:43 GMT
[Below is a joint discussion post from Steven and Ken]
Brief recap:
- Murphy describes a “classical view” of concepts (implicit in the works of past psychologists and certain philosophers), presents both theoretical and empirical arguments against this view, and describes how relatively newer theories of human categorization address these arguments. At a high-level, these newer proposals involve the replacement of hard, logical category boundaries with soft/fuzzy, probabilistic boundaries. These proposals also acknowledge the role of more complex, relational features (though, as Murphy notes, it is not clear that Hull intended to exclude these).
- Prototype view: our mental concepts are summary representations that are generated, in some way, from exposure to a distribution of previously-observed instances.
- Exemplar view: we store our mental concepts as distributions of previously-observed instances. Empirical findings that may have led some to adopt a prototype view ultimately are ultimately effects of retrieval/sampling from these distributions.
- Knowledge view (compatible with either of the two views above): semantic knowledge is highly structured (e.g. hierarchical). Learning is guided by intuitive theories of domains (such as physics and biology) -- prior conceptual knowledge about a domain both guides and may be modified by future learning within that domain (e.g. when one has knowledge about organisms in general, and is tasked with learning and making inferences about a new type of organism).
Discussion questions (feel free to answer your favorite subset of the questions below, and to introduce new questions into the conversation):
1) First, a broad question: do you have any qualms with Murphy’s theoretical and/or empirical arguments against the "classical view"? For example, you might take issue with definitions in Murphy’s presentation of theoretical arguments, or highlight potential confounds in some of the empirical studies he’s described.
1.a) Otherwise, do you have questions (or snark to share) about any theories of human categorization presented in Chapter 3?
2) When you first learned about the classical view of how humans represent concepts mentally, did this view strike you as accurate and/or intuitive? Were you surprised to learn that the classical view was so open to empirical shortcomings?
2.a) If it’s likely that users share this operationalization of conceptualization, than what implications are there for the design of HCI systems?
3) When it comes to categorization, clustering, and concept learning, where do you think AI and machine learning currently falls furthest from human abilities (and/or vice-versa)? Why might this be the case?
3.a) And (how) might the study of human concept learning, representation, and use inform the development of AI and machine learning (and/or vice-versa)?
4) Can you propose hybrids of the prototype, exemplar, and/or ‘knowledge’ (i.e. ‘theory theory’) views of human categorization? Feel free to reference existing models or theoretical frameworks.
5) How might certain features of human concept learning and representation (e.g. typicality effects in category learning, and in retrieval of members of a category) be adaptive or maladaptive in certain contexts (give examples)?
5.a) How might you design technologies that either leverage these features or minimize their undesirable effects?
6) Niki often incorporates ‘list building’ exercises into his lectures (e.g. asking the class to collectively generate a set of features that partially define ‘good research’ or ‘bad research’).
6.a) Do you have any thoughts/questions about the nature of the lists we’re likely to generate during these exercises, after reading Murphy’s book chapters (and perhaps other related literature)?
6.b) On that note: can you imagine interesting variants of Niki’s ‘list building tasks? What effects might these variants have on the process of list production and/or on the nature of the lists produced?
7) Can you recall a category learning task (from daily life) that was consciously challenging for you? Why do you think it was challenging? What feedback did you have from the external environment as you attempted to learn the category?
7.a) Via introspection, how do you think this category is represented in your mind (either with reference to the theories presented in Chapter 3, or moving entirely outside of these)?
8) A very open-ended question: On page 26 (Chapter 2), Murphy writes: “One reason for the popularity of the classical view has been its ties to traditional logic (Inhelder and Piaget, 1964)...”.
In modern research on human concept learning, it seems that probabilistic (especially Bayesian) generalizations of these logical models are popularly used in their place -- mirroring shifts in other disciplines such as AI and philosophy. This may be an instance of a (potentially) interesting general trend: theories of human cognition following and being shaped by tools/technologies (another recent example: modeling human memory with Google’s PageRank [https://cocosci.berkeley.edu/tom/papers/google.pdf]). Thoughts?
|
|
|
|
Post by mrivera on Mar 27, 2016 17:14:01 GMT
RE: (1) do you have any qualms with Murphy’s theoretical and/or empirical arguments against the "classical view"?
From my perspective, the idea of a one-descriptor summing up a category doesn't seem all that far-fetched or limited. Consider the statement all "dogs have 4 legs". It is true that generally dogs do have 4 legs. It is also true that a dog may have suffer an accident later in life that causes it to lose a single leg. These descriptions are lacking clarification. In fact an animal having 3 legs as a result of losing one doesn't disqualify it as being a dog because at some point, the said animal did have 4 legs. Furthermore from a biological standpoint, 4 legs (AFAIK) is an aspect of a dog's genotype. We could specify placement in a category on the basis of biological composition (at least for living things - either it fits a DNA profile or not). Generally our perception restricts our ability to define categories for things (we can see physical aspects, hear how something sounds, but we can't see chemical composition, DNA, at least not without extra work). Are there other ways we can test for categorization without strictly what we see and feel from direct interaction with a object/thing/being? (Do I have to see a photo of a dog, or interact with one to understand or categorize it as a dog)? I'm just spitballing
¯\_(ツ)_/¯.
|
|
|
|
Post by jseering on Mar 27, 2016 21:10:08 GMT
I buy the argument made by Wittgenstein; certain properties of categories can be articulated, but categories are inherently fuzzy. This is not surprising. Most of the categories that we use commonly in language have been around for hundreds if not thousands of years, and have been contorted and twisted and forced to adapt across different cultures and eras etc etc. Categories developed because they were useful for communication, not because they had any inherent validity. This, of course, causes a lot of problems in areas where we're forced to rely on common understandings of categories for high-stakes interactions. As the second chapter mentions, law is one of those areas; a typical contract is enormously bloated with complicated language designed to cover a broad array of contingencies, and even then we have multiple levels of courts and a long history of legal scholarship that has tried to define this language.
It's important that we recognize the fuzziness of categories and how their definitions are based in cultural context. We can't communicate without relying on some common assumptions about the meanings of categories, but getting too stuck on exact meanings can be problematic. See for example recent debates about the exact nature of gender/sex categories.
|
|
aato
New Member
Posts: 16 
|
Post by aato on Mar 27, 2016 23:36:18 GMT
I'm diving kind of in-between question 8 and my own question that came up in this reading. Both the prototype model and the exemplar model prompted me to start thinking more about how technology both supports faulty human categorization and may also contribute to it. An HCI researcher friend of mine has been studying bias on the internet. One of the main examples she uses to demonstrate bias is that on a normal Google Search for "grandma" the entire first page is of older white women with wrinkles and relatively short gray or white hair and almost all with glasses. This 'prototype' of what a grandma is is relatively pervasive in pop culture even though we know that grandmothers span a large variety of age ranges and races. Google images may both be reflecting what people in America/Western culture think of as a typical grandma, but is also perpetuating that prototype to everyone in the world who uses Google.
So in reference to the paper that Steven and Ken are citing in question 8 - PageRank may be good at predicting human response to fluency tasks both because it is good at modeling human behavior and cognition around these types of recall/fluency tasks, but the prevalence of PageRank may also mean that the results of these searches are constantly being renewed and strengthened. These also raises questions about generative categorization as Ken and Steven also mention in reference to Niki's list-building exercises vs. recognition tasks. If you ask someone to describe something more complex like a 'grandma' would you still come up with similar results to Google images?
|
|
|
|
Post by Felicia Ng on Mar 28, 2016 15:19:31 GMT
Kind of related to Ken's Question #3:
One thing that struck me as I was reading through these elaborate descriptions of different theoretical, empirical, and mathematical models is the discrepancy between how fast/easy it is for the human brain to make subconscious categorization judgments and how difficult/confusing/messy it is to develop explicit, quantitative, and algorithmic ways to imitate what the human brain can automatically do in milliseconds. This leads me to question whether machine learning and AI are even the best tools that we should be working towards developing to make categorization judgments. With the rise of crowdsourcing platforms, I wonder if we should be taking advantage of the unsurpassable human ability to make categorization judgments and capitalize on that instead of trying to replace it with computers. Going back to the questions of "What can humans do that computers can't do?" vs. "What can computers do that humans can't do?" I think we should be thinking about how to best divide up the tasks in the world in such a way that the unique strengths of humans and the unique strengths of computers can complement each other in an optimal/synergistic manner, rather than replace or imitate each other per se. Maybe categorization should be kept in the domain of human tasks? (Though I guess a counterargument is that it's not scalable or efficient if we solely depend on humans to make all categorization judgments.)
|
|
|
|
Post by Anna on Mar 28, 2016 21:31:58 GMT
First to piggyback off Alexandra's post and partially respond to Ken's #6: when we're coming up with lists of things that fit into a category, we're likely to come up with 'typical' examples. As Murphy points out in the chapter, however, what is 'typical' is not clear. Let's say we're making a list of features of grandmothers. We might base this on our personal experiences (our own grandmothers), but I wonder if it may actually take less cognitive effort for us to pull from cultural prototypes Alexandra discussed, even if they don't fit our own experiences. In a related example, if someone said make an exhaustive list of features that define beauty, I am likely to just pull from what popular culture tells us is beautiful, even if my own concept of what defines beauty is counter to these cultural ideals. So does the act of categorizing implicitly encourage conformity to cultural norms? If we ask people to engage in a categorizing task, absence of any priming, will they be likely to give greater weight to cultural norms rather than their own personal experiences and opinions?
I also want to respond to Ken's #7. So, I wouldn't say it was extremely hard for me, but as a toddler, I had trouble distinguishing between adult genders, sometimes to the embarrassment of my mother. As I grew older, this became a relatively easy task. But now as an adult, I'm more aware of fluid gender identities/gender identities that don't fit into binary categories, gender identities that are not always accepted in traditional value systems, etc. This seems to fit in with Murphy's example of metallurgists-- sometimes with more knowledge, categories can grow fuzzier rather than clearer.
|
|
mkery
New Member
Posts: 22
|
Post by mkery on Mar 28, 2016 23:30:56 GMT
Responding to #3, I wonder how AI/ML fair at tasks requiring creative combination of concepts. A simple example may be “dog-cow”: can you draw a dog-cow such that it is recognizable to others as a dog-cow? This may require making sure important features of a prototype dog (e.g. a long tail) and important features of a prototype cow (e.g. black and white spots) are present. How do you choose features of these animals, however, such that a dog-cow cannot be mistaken for a cat-cow (perhaps a thin tale) or an obese dalmatian?
Humans are very capable in creative tasks of recombining concepts. I’m curious how the theories in this reading apply to artificial concept-creation that happens in any creative work with fiction or art. I worked briefly as an undergrad in a lab that attempted to apply AI to make procedurally generated storylines. A possible application of this work is creating video games with infinite branching depending on the choices your character makes. (cool.) However, I was surprised how extremely limited current AI is in this task. Simply combining events such that the story is plausible to a human is a very difficult and very time-consuming search task through thousands of possibilities. If you give a human a highly complex task “write a new story that has the same concept as Harry Potter”, humans are highly capable of making logical, if truly terrible, fan fiction. An AI agent will struggle to tell you how Harry would react if he confronted Voldemort on a beach in Auruba: this requires a concept of each of the characters, their motivations, their past interactions, a concept of a beach, a concept of a beach in relation to the rest of the story, ect. Algorithms are logical within finite rules but struggle to produce the complex logic humans create out of relationships.
Scaling back to simpler tasks, how are AI/ML at concept creativity?
|
|
|
|
Post by francesx on Mar 29, 2016 1:12:33 GMT
Let me start of by saying that according to NASA Pluto is a dwarf planet (http://www.nasa.gov/audience/forstudents/k-4/stories/nasa-knows/what-is-pluto-k4.html). I agree with Mike's point that for living things we can use DNA to "define" them. Even though I am not totally positive, I remember studying how much argument(s) in biology/science went into "how to classify living things" starting off with what are "living things" (cats? trees? sea squirts?). My point being, even using the DNA is not always easy.
In addition, this came up during the discussion for the paper Michael Madaio and I will cover: categorization, classification or clustering is based on "similarities" among different "things", and differences these things have with other categories. Which for me raises a question: why do we need to classify, and moreover as human beings why do we "always" try to classify anything into something? (Or do we?)
Second, I always thought that AI did a better job than humans in categorizing and clustering information (questions #3), at least bigger/greater quantities of information. I would like to say that AI can not make good judgment calls, based on logic, as we humans can do. But on the other hand, can we humans really do that? I agree that the study of human learning and representation, could not only help AI (by making better design and choices), but also us human, in doing a better job in learning and categorizing.
|
|
|
|
Post by mmadaio on Mar 29, 2016 1:34:44 GMT
To address Question 6... I've noticed this about the class activities (in this class and P&T), where a list of representative items is solicited for a given topic, and then these are either clustered and categorized, or weighted, or voted on. However, as has been discussed in the readings, there might be a more structured relationship between the features, or some may be mutually exclusive, or conditional, etc. One could imagine a similar activity to elicit spatially structured features (perhaps in a concept map, or so-called "mind-map", or some such other network based approach), or some more complex logical relationship between sets of features (X not Y; if A and B, then C; etc). But, I'm not sure this would be THAT much more pedagogically valuable to make it worth the extra effort. Possibly, and maybe for some topics more so than others.
To Ken and Stephen's last question about models of cognition following the tools and technologies we develop... you can look back at the history of writing about the mind and see metaphors for cognition drawn from our tools, from the steam governor used to explain feedback models of cognition to maintain our "mental equilibrium", to metaphors of our brains (and the universe) working like clockwork in the 13-1500's as clocks spread across Western Europe, and even from Homer, describing our thoughts flying like arrows, or like oarsmen on a ship. It's hard to say whether this is inherently good or bad, but it's difficult to imagine it any other way, so long as we recognize the limitations of the metaphors we use. (i.e. describing our brains as being "wired" for something implies a certain rigidity or consistency (or intentionality and finality of design) to their structure, and removes our agency over the re-structuring of our brain activity). Metaphors have consequences, ya'll.
|
|
|
|
Post by xiangchen on Mar 29, 2016 1:51:31 GMT
The 'counter-classical' view seems quite intuitive to me, especially given the examples (defining game, car seats-chair-furniture complication). Two thoughts from reading this. First, mathematics has found no problem whatsoever using strict and discrete definitions. Why? Is it because mathematical items are only theoretical and do not have the 'fuzziness' of real world items? Or is it easier, cognitively, to accept certain categorization in the math world where in the real world disputes are prone to happen? Second, discrete labeling has been quite successful in machine learning. Decision tree, for example, can nicely solve a complicated problem simply by creating hierarchical, discrete categories. We do need categories (the classically viewed one), don't we?
|
|
|
|
Post by xuwang on Mar 29, 2016 2:32:01 GMT
In response to #3, I think still a lot of classification task relies on human labeling in the first place. And for unsupervised learning/clustering, for example LDA topic modeling, it’s able to generate certain number of clusters of words, each cluster represents a topic, but still we’ll need to manually label each cluster to give it a meaning. I think machine learning models are able to find similar terms, but they are not good enough to give meaning to each category. And machine learning models may perform well for tasks that the distinction between categories is obvious, but may not be able to detect the subtleties if we have subcategories.
In response to #6, I think the approach Niki used is more of a bottom-up approach to let us brainstorm all the possibilities and later categorize (or may not categorize) them. I think an alternative way could be a more top-down approach to give certain big categories based on our previous understanding, and fit the rules each of us mentioned into these categories. For example for the question “what human can do while computers can’t do”, there can be categories of “cognitive” “emotional” “social”, etc., this may make information more structured, but could also make some rules hard to categorize.
|
|
judy
New Member
Posts: 22
|
Post by judy on Mar 29, 2016 2:37:58 GMT
Quick thought on lists in class: in my training as a teacher/facilitator, I was encouraged to create lists and mind maps when introducing a topic in order to "access the knowledge in the room." Then by writing up what we know (or think, or believe) about a topic we could start to group things together (in whatever way is appropriate...if it's a history lesson, maybe we group things chronologically; if it's a polarizing topic, maybe we identify the different arguments) and find the gaps in our collective understanding. This way it is the class that is making sense of what we know, and at least publicly acknowledging (if not coming to some consensus) about what we don't know. Then as we read/practice/learn more we can fill in our map of an area. Voting is less about reaching a consensus than forcing each student to think through each option themselves and form an opinion/hypothesis in their own mind, and then have a chance to see how other people have formed the same or different opinion.
On Alexandra's point about bias: I'm going to try to apply the problem of biased google search results to the Exemplar View. Let's see. If in my experience the grandma's I've seen are mostly old white ladies, then those are the only set of grandma's I have in my memory. So when meeting a woman for the first time, I wouldn't even consider whether or not to classify her as a grandma, unless she fit in with the set in memory. As I meet more people in the world (or consume more media), I commit a more diverse set of woman to the category of grandma, and so think might think to myself when I meet someone, "is this person a grandma?" So...if google returns images of old white ladies, then it's because it only "sees" images labeled as "grandma" that are old white ladies in rocking chairs. (This is getting obvious!) Then is that because the masses are overwhelmingly "seeing" (visiting pages with) images of old white grandmas? Who is looking at images of woman from other ethnic backgrounds? With other lifestyles? In this way, does the act of looking, of going to the web page, become a way of representing or of "being seen"? Is looking active? Is looking sometimes a political action (in the way that we sometimes think of consumerism as a political action)?
|
|
|
|
Post by julian on Mar 29, 2016 2:43:56 GMT
3) When it comes to categorization, clustering, and concept learning, where do you think AI and machine learning currently falls furthest from human abilities (and/or vice-versa)? Why might this be the case?
They should not be compared. On one hand categorization and clustering on ML are about finding automatically groups using a metric. Usually members of the same group minimize this metric. This could be seen, from the classical way of a concept, as a way to find the necessary and sufficient conditions to determine a concept or a category. But, there is nothing here related with understanding, concept learning or anything related. Although for us it may look like an algorithm categorized something in a humanly way, it just solved a mathematical problem.
Now contradicting myself, and ignoring the understanding part, ML and AI algorithms can effectively categorize high dimensional data, however it has its own shortcomings. For instance, when the number of variables is higher than the number of examples it becomes an under-constrained problem hence results are not stable. Even if there are enough examples, metrics like the euclidean distance become much less effective in high dimensional data. This basically means, that even for ML and Ai algorithms categorization is not a trivial task.
Now there are very important implications for HCi on the intelligibility of clustering and interpretability of the results. Even if a machine can categorize high dimensional data, that does not mean we can easily explain why a category or group is a group (Dah! the metric) and what are the important characteristics about it (this becomes a very difficult problem for even 25+ variables, how do you talk about this). In a project I'm currently working on, we experienced all of the above problems first hand.
Currently(and since always), it is very difficult to apply Ai and ML in healthcare because of the above reasons.
3.a) And (how) might the study of human concept learning, representation, and use inform the development of AI and machine learning (and/or vice-versa)?
I think it has already. The way clustering methods work in a way resembles the most basic way to categorize. This methods are also plaged with the same problems we have on the human side: Fuzzy definitions, categories that are dependent on the algorithm, the right number of categories is usually unknown and it is very hard to identify.
In tipicality and atipicality: It is pretty much the same on ML and Ai, atypical (less frequent) examples are more difficult to recognize as belonging to a category.
Now on the less classical (maybe they are now classical) theories:
The prototype view can be thought of as Kmeans, centroids are the summary representation and are in fact an average representation of all data points in the category.
Feature combinations can be found as Feature weighted Kmeans
The schemata does not have any equivalent that I'm aware of.
The Exemplar is basically K-nearest-neighbors.
|
|
|
|
Post by adamstankiewicz on Mar 29, 2016 3:04:47 GMT
My response is in response to the question of how the classical view of concepts is open to many empirical shortcomings and what implications there are for HCI systems (2 and 2a). In the classical view, concepts are mentally represented as explicit definitions, where something (e.g., a dog) either fits the definition of a concept or it doesn't (there's no middle ground). Murphy discusses how this leads to typical and atypical objects in a concept (e.g., good example vs. bad example, respectively). The typical items are considered cognitive reference points, mostly because they are more familiar with the expected object that matches a concepts definition. However, one way the classical view breaks down is with the notion of transitivity in which if an object fits a concept that is a subset of a higher-level concept, it is logical for it to be included in the higher-level concept as well. Despite this, empirical studies showed that this was not always the case.
I think this could have implications in social computing systems such as recommender systems. If you consider a movie recommender (e.g., Netflix), the system might recommend movies that are similar to ones that you've watched before in similar categories or genres of film. However, depending on how the system internally classified movies, it may think a movie is in a sub-genre of a movie genre you seem to like, yet you really don't care for the sub-genre for whatever reason. This maps onto the transitivity issue discussed by Murphy in which the classical view of concepts breaks down.
|
|
k
New Member
Posts: 9 
|
Post by k on Mar 29, 2016 3:06:22 GMT
Murphy argues that the classical view is flawed by pointing to empirical studies that undermine the classical view's depiction of categorization. Yet, why should we think this an appropriate line of argumentation? Murphy ascribes the classical view to philosophers early on in the paper, yet, looks to the work of psychologists to undermine it. This confuses these disciplines and does not address the methodological differences between the two. Murphy argues that the classical view is committed to revising/rejecting conceptions of categorization in the face of counterexamples. Yet, why should we think the empirical studies of psychologists provide compelling counterexamples?
|
|
