Psychology 120A

One piece of evidence in favor of this theory is...

Murphy and Allopena (1994) found participants have difficulty learning about things that do not "make sense" - When we learn about categories we try to make meaningful connections from our past knowledge. - We rely on categories to teach us about the world and use our knowledge about the world to help explain category membership. - They did something very similar to the random dot pattern experiment and the diggers vs. builders experiment. Except, here they were having participants learn how to categorize buildings. - The clever manipulation: sometimes buildings had sets of features that made sense together and other times the features didn't really make sense together. In one case, the one where it made sense could be that a building has really thick windows, it's red, divers live in it, it's underwater, and you get there by submarine. These might be sort of the features of one type of building and participants could learn this type of building pretty well. Other times however, the features wouldn't really make sense. For example, one building in the category that doesn't make sense could have steel windows, be purple, farmers live in it, it's in the desert and you get there by submarine. In this case, it doesn't really make sense. Participants had a hard time learning this type of category that didn't make sense. So this really shows that it's not just a list of features, it's more a theory of what does it take to be this kind or that kind of category.

Neural Network Models

Unlike semantic network models, artificial neural networks (ANNs) don't store knowledge explicitly in nodes. Instead, knowledge is contained in the distribution of weights between the connected nodes.

What are the two limitation to Prototype Theory?

1. Context Effects 2. Typicality Effects

Commonsense Knowledge Problem

Humans have a lot more knowledge than computer scientists ever realized.

Method of Repeated Reproduction

(Bartlett) in this method participants are shown a stimulus then asked to reproduce it from memory, and again, and again, etc.. Importantly participants are only shown the original stimulus once: each reproduction has to be based on memory. You can see in Figure 9.14 that the reproductions become less similar to the original stimulus with each attempt. What you can also see is that the reproductions begin to look more and more like a familiar object. What started out as an abstract image, ended up becoming a human face. Over time, details are lost from memory but we can use the information from our schemata to help guide memory retrieval.

Typicality Effects

- Behavior directed differently toward typical items compared to atypical ones. - When listing category members, we name typical members first. - Faster to put typical members into categories than atypical ones. - Typical category members show priming effects that atypical members do not. - Lexical Decision Task

Random Dot Pattern Experiment (Posner & Keele, 1968)

- Four random dot patterns serve as category prototypes - Participants see 12 distortions of each prototype - Learn to categorize patterns with feedback Then we will test categorization accuracy for - Old distortions (exemplars) of prototype: the exemplars that they were just trained on. - New Distortions (exemplars) of prototype: some you've never seen before - New Distortions (exemplars) further removed from prototype - The unseen prototypes themselves (remember they've never the prototypes before. Results: when people see the old distortions, people do pretty well. The new distortions, people do not do good. However, with the actual prototype, they are very good at recognizing them. This suggests that people are extracting the prototype because when they saw the prototype, which they have never seen before, it is as if they have already seen it.

Limitation to the Prototype Theory: Context Effects

- The individual exemplars that we are exposed to in our lives do play a role in creating what is typical and what is not. So it cannot be that we see the exemplars and we derive from them a prototype and forget the exemplars. - Typicality should depend on the number of shared features between category members. It doesn't. - Typicality depends on context. - Prototype theory doesn't explain how characteristic features can differ between contexts.

Prototype

- The most typical member of the category - An abstraction that has all the characteristic features of a category. - Individual exemplars do not play a role in categorization.

Knowledge-Based Approach to Categorization

- We rely on our broad knowledge base and intuitive theories about how the world works to explain the reasons for category membership. - The idea at its core is that... Psychological Essentialism: the idea is that all category members possess a fundamental essence that is unique to that category and determines membership. Example: Dogs are "doggy", birds are "birdy", and fruits are "fruity" Example: You're studying, it's late at night, and suddenly you hear some noise coming from the alley so you look outside and there's a person who is intoxicated, maybe they have a bootle, they're wobbly, rolling around a puddle, they have a jersey etc.. So you look at that person and say, "Wow, that person is drunk''. So, you categorize that person as drunk. Of course, it could have been that in your mind you have a list of things that it takes to be drunk smelly breath, wobbly walk, rolling in puddles, singing the national anthem very poorly, there's a long list of things and the more of these that one person has the more likely they are to being drunk. However, according to the theory based approach, it's not about a list of features that you have or you don't have that make your less or more likely to be part of one category or another. It's just that you have a theory of what it takes, what it means to be drunk. At its essence being drunk is really a theory of impaired decision making, so all of these features, all these things that you observe, make sense if a person is impaired because they drank. - It's not about list and features, it's about theories of how things work and that is how you place items into categories.

Basic Level Categories

- When we categorize things we tend to categorize them at their basic level also referred to as their natural level. (Think of the cow he showed us and how he predicted that a majority of people would say "Cow".) - The level at which we talk about things, provides just about the right amount of information about the category that can be informative when you're talking to someone else and can be used to distinguish members from members of other categories. - Informative and Distinctive

Concept

- a mental representation of objects, ideas, or events. - The fundamental unit of symbolic knowledge in your mind.

Category

- a set of items that are perceptually, biologically, or functionally similar -Items in a category are considered equivalent such that one item can be exchanged for another with little consequences. -For example, if you want to sit down, any chair will do. There are different types of chairs, but for the most part, one chair can be exchanged for another and you will still be able to sit in it.

Classical View of Categorization

- categories are defined by sets of features that are both necessary and sufficient for category membership - The mental representation of a category is abstract (i.e., there is no information about individual exemplars) Example: - Category: Dogs - Defining Features: Four Legs, Has Fur, Barks - Is a three-legged dog still a dog? Is a dog that can't bark still a dog? Yes, they are still dogs!

Exemplars

- individual items in a category (different chairs in the category "chair"). - Within the category of pets, we have exemplars of dogs, cats, birds. - Pets is an exemplar of the category of animals

Subordinate Categories

- more precise categories - When we talk about items at their subordinate level, they tend to share a lot of features with other items.

Exemplar Theory of Categorization

- proposes that we store in memory examples of items we have encountered in the past. - Categorization occurs by comparing new items to the ones you have in memory and looking for similarity between their characteristic features. - This theory does say that if needed you can also extract a prototype, but not necessarily. The important point is that you bring with you your experience, your history of the exemplars you've seen, which explains why we can have context effects.

Prototype theory states that individual items can belong to multiple levels or Hierarchies of categories.

1. Superordinate Categories 2. Basic Level Categories 3. Subordinate Categories

Limitations to Both Theories

1. Typicality Ratings - Armstrong, Gleitman, and Gleitman (1983) did an experiment in which they asked participants first, to categorize numbers as even or odd. Participants have no problem telling apart even and odd numbers. Then they asked how typical is 3 of an odd number and how typical 447 etc. etc. It turns out that systematically people view numbers like 3 to be much more typical of being odd than 447. This makes no sense because the category of oddness does not have fuzzy boundaries, it's very well defined and yet we behave as if it had fuzzy boundaries and therefore we have these judgments of typicality. - Is typicality actually a result of experimental design, rather than the fact that categories have fuzzy boundaries. 2. Lack of a Hard Boundary Between Categories - Prototype/Exemplar provide a continuous typicality rating, yet our categorization intuitions are all-or-none. Yet, our intuitions don't really work this way. Something like a tomato is either a fruit or it's a vegetable not 51% fruit & 49% vegetable, therefore I'll go with fruit. That's just not how our minds work, when we see something we just categorize it, you're in or out. 3. Theories are Similarity-Based - Vague - What matters with regards to similarity? - Variable (e.g., Context-Dependent) - So our judgements on similarity depend on the context. - It's difficult to establish firmly what we mean by similarities, similarity is a very vague concept, it's hard to tell what is the frame of reference, what matters with respect to similarity. Yet, when people ask us we seem to have very clear ideas as to what counts for similarity however, neither theory accommodates this kind of thing.

Do we really store all the items we have previously encountered?

Allen and Brooks (1991) trained participants to identify drawings of creatures as either "Diggers" or "Builders" using a rule based on physical features. The rule was to be a builder you need to have at least two of the following three features, long legs, angular body, or spots. If you have any two of these you're a builder or else you're a digger. They were first trained on a number of exemplars Then they were shown new exemplars and asked to categorize them if they were builders or diggers. The question is how often will participants mistake category C for being diggers versus category D for being diggers. If we use the rule, then participants should equally look at category D and say they are obviously builders (because they have long legs and spots). However, it turns out that new creatures that were physically similar to the exemplars that participants had seen during training (Category C) they only mistaken them for diggers 20% of the time. However, category D even though they are builders according to the rule, but do not physically look like builders, were mistaken for diggers up to 45% of the time.

Reconstructive

Bartlett demonstrated that memory is reconstructive. Instead of retrieving an exact copy of an event from memory, we rely on our past knowledge and experience to help us reconstruct memory the best we can.

Spreading activation model of semantic memory

Collins and Loftus (1975) An alternative semantic network model was proposed that could account for typicality effects Definition: nodes are connected to each other via semantic relatedness (and not hierarchical structure). - The nodes that are closer to each other are nodes that are semantically related. So the idea is that the hierarchy is gone, nodes are close to each other in proximity to each other in relation to how semantically. - The more semantically similar nodes are the more connections there will be between the two of them and the short the distance between the two of them in the network. Thus, when one node becomes active, activation spreads to all the connected nodes. - There is no hierarchy, concepts are organized by their semantic similarity. - The more similar the concepts, the more connections between them and the shorter the distance between them - When one node becomes active, activation spreads to all connected nodes. - The farther the activation must travel, the longer it takes and the weaker it becomes.

Semantic Network Models

Collins and Quillian (1969) - knowledge is stored as concepts within a network of interconnected units called 'nodes'.

Hub-and-Spoke Model:

HUB: According to this model, generalized and abstract semantic knowledge is stored in a semantic memory hub in the ATL. This is where your general knowledge of "apple" would be stored, including all the places you could find an apple and all its different uses. SPOKE: context-dependent and modality-specific detail about items is stored in "spokes" that are distributed across the cortex. What different apples look like would be stored in visual processing brain areas, what an apple tastes like would be stored in taste perception cortical areas, and how to hold an apple would be stored in motor cortical areas.

Limitation to the Prototype Theory: Typicality Effects

How do we account for atypical category members? Example: A penguin, how could a penguin be considered a bird if we would compare it to prototypes. Very few people grow up with penguins around all the time as birds, yet, we recognize penguins as birds. So it cannot be that all we do is extract from the exemplars that we find, a prototype and then just use that in order to categorize new exemplars.

Sentence Verification Task

Participants were faster at responding to,"A Canary is a Bird" compared to Canary is an Animal. Rips, Shoben and Smith (1973) - Participants are faster at verifying "a dog is an animal" compared to verifying "a dog is a mammal" - "Robin is a bird" faster than "a chicken is a bird" which is odd because both robin and chicken should be daughter nodes of the node bird. - So this model cannot quite account for typicality effects.

Is there any evidence that we actually do this? Do we abstract categories?

Random Dot Pattern Experiment (Posner & Keele, 1968)

Typicality Ratings

Ratings of items based on how good of an example of a category an item is.

Symbol Grounding Problem:

The problem is that any symbol system can only replace one symbol with another one, and this process could continue infinitely. - There needs to be some way to connect this symbolic representation to the real world; in other words, symbols need to be grounded. -Example: A better way to explain what an apple is would be to show the alien an apple and have them take a bite

Semantic Dementia

a progressive neurodegenerative disease characterized by an inability to name objects, but importantly, this deficit is not primarily a language deficit or a perceptual deficit. That is, patients are unable to name objects presented visually, verbally or by touch because they have a deficit with the knowledge itself, not with processing input from one of the senses Semantic dementia is associated with the degeneration of neurons in the anterior temporal lobe

Schemata (plural of schema)

a schema is our organized knowledge-base about a particular topic. It includes everything we know about a particular thing, event, person, or situation. - Example: University Schema Includes everything we know about what a university looks like, what you do there, and what people you would find there. - The concept of a schema is broad and may seem vague, but that is because the knowledge within a schema is broad and not necessarily well-defined.

Family Resemblance

another way of thinking about using characteristic features to determine category membership is to think of category members as having family resemblance.

Ad-Hoc Categories:

are an example of the flexibility of categories. - Barsalou (1993) stated that categories are flexible depending on the context. - These categories are not necessarily stored in memory, but they come together as categories only as they relate to a specific goal. Example: Can you guess what category these exemplars are part of? Move to the remote regions of Wyoming Change your identity Sail around the world Go to Mexico What category are they a part of? (Things to do to avoid getting killed by the Mafia).

Graceful Degradation

because knowledge is stored as a pattern of activity across a large number of units, connectionist networks can withstand some loss of units with limited negative effects. It is commonly observed with damage to the human brain. A network (or brain) doesn't lose all function as a result of restricted damage, although some limited deficits do occur. For example: patients with brain damage as a result of Herpes Simplex Virus Encephalitis often display Category-Specific Deficits of semantic knowledge. These patients frequently lose their knowledge of living things but maintain their knowledge of non-living things.

Superordinate Categories

distinctive, but not too informative.

Characteristic Features

features that are likely to belong to a category member but are not required for category membership

Semantic Memory

is our long-term memory for information that has lost its sensory details.

Cognitive Economy

store a property only once at the highest level

Property Inheritance

subordinate categories inherit the properties of connected superordinate categories. - The idea is that each node inherits the properties of it's mother node. - Example: Bird- has wings, can fly, has feathers, but also inherits all the properties of Animal-, so Bird- also has skin, can move around, eats, and breathes. - This leads to the idea of cognitive efficiency. We store concepts only once inside this network, and we store them at the top level.

Meyer and Schvaneveldt (1971)

used a Lexical Decision Task to demonstrate semantic priming between two related words. - Participants were faster to respond to the word "butter" if it was presented with the word "bread" compared to if it was presented with the word nurse. - The spreading activation model explains typicality effects because typical exemplars are semantically similar to each other, close, and therefore activation will spread quickly between them. - Atypical exemplars are farther apart from other category members because they are less semantically similar, so activation will take longer to travel between those concepts.

Prototype Theory of Categorization

we categorize items using characteristic features to compare to a prototype stored in memory

Black Box Problem of Neural Network Models

while neural networks offer some advantages in representing knowledge, they are also notoriously difficult to explain or interpret once they have been trained. While we can observe the responses of a neural network to a specific input, it is very difficult to determine why it made the response that it did because the information is represented in the values of distributed weights, not meaningful semantic units