L&M Concept Learning PDF

Concept Learning (Lecture 13 & 14) ​ Semantic memory ○​ Semantic knowledge allows us to predict, understand, and respond to new experiences ○​ Semantic knowledge often arises from common elements across multiple (episodic) experiences ○​ Scientists h...

Concept Learning (Lecture 13 & 14) ​ Semantic memory ○​ Semantic knowledge allows us to predict, understand, and respond to new experiences ○​ Semantic knowledge often arises from common elements across multiple (episodic) experiences ○​ Scientists have formal methods to organize knowledge about how the world works ○​ How is semantic, declarative memory organized by the mind? ​ Must support retrieval of facts we learned ​ Must support reasoning about about relationships between those facts ​ Must allow us to make inferences about new experiences ​ Semantic networks and spreading activation ○​ The amount of time to verify a sentence depends on distance in semantic network ○​ Semantic knowledge is organized in hierarchical networks ​ Distance between concepts (nodes) in the network represents their relatedness ○​ Retrieval occurs via spreading activation through the network ○​ Semantic (or category) fluency test: write down as many items from a category as possible within a short period of time ​ Order of retrieval reflects semantic relatedness between items ○​ Based on responses to semantic fluency, we can reconstruct how animal concepts are organized in the mind ​ Size: how often people generated the animal ​ Distance: how often two connected concepts (nodes) were generated one after another ​ Priming in semantic memory ○​ Priming: exposure to a stimulus influences the response to a subsequent stimulus, often in the absence of explicit awareness ○​ Typical procedure: ​ Present a prime stimulus (usually during unrelated task) ​ The prime activates an internal representation in memory ​ Activation spreads to other, related concepts in memory ​ Activated representations affect behavior ​ Recall: distributed representations ○​ Mental representation of a visual object (like a face) involves layers of increasingly complex or abstract features as information flows through the visual system ​ Word recognition ○​ Words are objects composed of a fixed set of features (letters)  ○​ Comprehending text requires you to recognize words and retrieve their meaning from semantic memory ○​ Words can be recognized relatively easy even if many features (letters) are in scrambled order ○​ Interactive activation theory of word recognition ​ Words are represented as combinations of letters, which are represented as combinations of line segments (features) ​ Activations spreads through the network in both directions so that representations of words and letters interact ​ Recognizing words in context ○​ Word superiority effect: individual letters are recognized faster and more accurately when presented in the context of a real world ​ According to Interactive Activation theory, activation at the word level spreads to the letter level, making retrieval faster ​ Same effect doesn’t happen when letter is presented alone or in context of an unfamiliar nonsense word ○​ Interactive activation allows us to fill in the gaps when input is ambiguous or incomplete (example of pattern of completion) ​ Summary ○​ Conceptual knowledge is organized in hierarchical associative networks ○​ Retrieval of semantic memory occurs via spreading activation through those networks ○​ According to interaction activation theory, representations of lower level features and higher level concepts interact ○​ Word superiority effect: activation of word in memory facilitates detection of its features (individual level) ○​ Activation of higher level concepts allows us to generalize in conditions where input is scrambled, distorted, or dissimilar to prototypical experiences ​ Semantic memory ○​ Episodic memory ​ Organized by spatial and temporal context ○​ Semantic memory ​ De-contextualized ​ Organized by meaning ​ Generalizable knowledge ○​ Concept learning and generalization ​ Concept: a psychological representation of a category of objects, events, or people in the world ​ Category learning: the process by which animals and human learn to classify stimuli into categories  ​ Generalization: the transfer of past learning to new situations and problems ​ Predicting similar consequences based on perceived similarity to past experiences ○​ Generalization gradient ​ Pigeons trained to peck at yellow light to get reward, then tested on response to range of colors ​ Generalization gradient: graph showing how physical change in stimuli correspond to changes in behavioral responses ​ Not a result of confusing different stimuli or being unable to perceive differences ​ Strength of response represents prediction that similar stimuli will lead to the same outcome ○​ Generalization in associative networks ​ What does generalization tell us about how different stimuli are represented in memory? ​ If different stimuli are represented independently, no generalization would occur ​ When the output node is activated, a response is generated ​ An internal layer contains nodes that receive input from multiple, similar stimuli ​ Input nodes are activated when corresponding stimulus is presented ​ Distributed representation: a representation in which information is coded as a pattern of activation distributed across many different nodes ​ To be able to generalize, we can’t treat every experience as a one-time, unique event, must be able to recognize shared elements across experiences ​ Perceived similarity between different stimuli depends on whether they have overlapping representations in memory ○​ Category learning ​ Classical view of category learning: learn the set of features that define members of the category ​ Categorization involves comparing a new object to the list of features necessary for category membership ○​ Category learning: prototype theory ​ For many natural concepts, there is no simple set of conditions to determine membership (contradicting the classical view)  ​ Category membership is graded, not all-or-nothing (we view some items to be better examples of a category than others) ​ Family resemblance property: category members tend to have overlapping sets of features in common ​ Prototype: the central tendency or idealized version of a concept or category ​ According to the prototype theory of category learning, a learned prototype combines the most common features across experiences examples of the category ​ Categorization of a new stimulus occurs by evaluating its similarity to the stored prototype ​ Features can be physical or non-physical ​ Typicality effect: objects that are more similar to the prototype are judges as better, or more typical, examples of the category (and are classified more quickly) ​ Ambiguous, difficult to classify objects tend to be equidistant to prototypes of multiple categories ​ Solso and McCarthy experiment ​ Participants studied a set of faces for 10 seconds each ​ Studied faces were all highly similar to a prototype face which was never shown ​ After a delay, presented a new set of items that included the unseen prototype, the previously seen faces, and new, dissimilar faces ​ Asked people whether they recognize each test face (and how confident they were) ​ Participants confidently recognized the prototype face even though it had never been shown ​ Among new (unstudied) faces, confidence was highest for faces with higher proportion of shared features ○​ Configural representations in face recognition ​ Inversion effect: loss of ability to recognize abnormal features when upside-down (changing the configuration) ​ People can recognize faces from caricatures more quickly than from real photographs ​ Caricatures emphasize configurations of features that are present in representation of the person (and unlikely in memories of other faces) ​ Summary  ​ We use perceived similarity to generalize from existing conceptual knowledge to new events (generalization gradient) ○​ Generalization provides evidence of distributed representations which allow us to recognize shared features ​ According to prototype theory, memory for each concept is stored as a prototype that averages across experience ○​ New objects/events are classified by comparing distances to stored prototypes ​ Memory of a concept incorporates not only the set of features, but also how those features are arranged (configural representations) ​ Discrimination learning ​ Generalization: the transfer of past learning to new situations and problems ○​ Predicting similar consequences based on perceived similarity to past experiences ​ Discrimination learning: the process by which animals or people learn to respond to differently to different stimuli ​ Discrimination learning is important for learning how natural categories should be further divided in meaningful ways (based on differences in experienced outcomes) ​ Through experience, you may learn that category members are associated with different outcomes based on one feature dimension ​ Pigeons were trained to peck to get reward based on frequency of a one ​ Two conditions ○​ Standard training: responding to a single tone is always reinforced ○​ Discrimination training: responding to a 1000-Hz tone is always reinforced; responding to a 950-Hz tone is never reinforced ​ After training, given a generalization test: play a wide range of frequencies and measure whether pigeon responds by pecking ​ After discrimination training, pigeons learned to treat 950-Hz and 1000-Hz differently ​ Discrimination training led to less generalization overall compared to standard training ○​ Summary ​ We use perceived similarity to generalize from existing conceptual knowledge to new events (generalization gradients) ​ Generalization based on distributed representations in memory which allow us to recognize shared features ​ But studies of discrimination learning show that generalization is not determined solely by perceived similarity ​ Quickly adjust generalization gradients in response to differing outcomes ​ Discrimination training tends to produce more narrow generalization gradients, in effect signaling that the world is more complex than it appears based on similarity alone