Nosofsky (1984). Choice, similarity, and the context theory of classification. Journal of Experimental Psychology: Learning, Memory, and Cognition, 10, 104-0114.

Nosofsky, R. M., & Zaki, S.R. (1998). Dissociations between categorization and recognition in amnesic and normal individuals: An exemplar-based interpretation. Psychological Science, 9, 247-255.

Goldstone, R. L., & Barsalou, L.W. (1998). Reuniting perception and conception. Cognition, 65, 231 - 262.

Wisniewski, E.J., & Bassok, M. (1999). What makes a man similar to a tie? Stimulus compatibility with comparison and integration. Cognitive Psychology, 39, 208-238.

Common Threads, Reactions, and Future Research

 According to exemplar models of classification, people store individual exemplars in memory and classify new objects based on their similarity to stored exemplars. The first question was: Why would we do that? General discussion pointed to the answer that it assists us in classifying objects into categories. For example, if asked which is more likely to sing, a small or large bird, the more likely answer is large bird because retrieving examples from memory does not bring to mind any large singing birds. The idea is that examples that are most similar to the item to be classified are categorized quickly.  Exemplar models allow for predictions and inferences.

Editor: Lesley

There is no list of jointly necessary and sufficient features. Some exemplars are better than others. There is a graded category membership. We know this because we have a mental explanation of things that goes beyond similarity. We consider  features differently, and some categories are more salient. The degree to which features are related to each other is important. Also, defining features are salient even when they are not frequently encountered. So, although a frequently perceived feature helps classify objects quickly, it is not enough. We have to consider more than similarity and also frequency of association of a feature to other exemplars in the category. We have to be able to make inductions that are neither too broad, or too narrow. We have seen how we use similarity in sentence verification. We verify sentences such as, A robin is a bird, much faster than sentences such as, An ostrich is a bird.

We discussed homeopathy and contagion. We brainstormed examples of homeopathy which is when the cure and cause appear to be similar to the symptoms. We could not think of any concrete examples in our culture, but we did discuss Rozin, who presented a colloquium at KSU last year on Disgust. For example, people will reject chocolate if it is shaped into a disgusting object. Contagion is basically a cause and effect paradigm. The germ theory was the obvious example. Originally, people could not believe that devastating diseases were caused by such tiny germs. It is unlikely, however, that they did not recognize that some form of contact with infected people resulted in many diseases.

The argument against the exemplar-based model, that we were concerned about, was that it did not adequately explain how we could classify novel objects. We need a method of accommodating to novel objects in order to learn them. Clearly, having an exemplar-based system enabled us to benefit from experience that we could generalize to new situations. However, if we did not have a system that enabled us to create new exemplars, we would be unable to learn. The big question was how are concepts structured so that if you learn about one thing, you can generalize to others? One of the ways of generalizing is based on similarities. Objects are not just similar in perception. We have rules and criteria that have to be met, for example, a prime number has very precise rules that we can state easily once we have learned them. Exemplar and Prototype models are actually equivalent. The problem was in defining similarity and this is where the articles offered different approaches.

Context theory was interpreted in terms of choice theory and similarity to exemplars. The relationship between identification and classification was explained. An important rule to consider in the description of similarity is the response-ratio rule. That is, given an object X, how do you classify it? Context theory assumes people classify stimulus based on their similarity to stored category exemplars and the probability of classifying a stimulus as a member of a particular category is a ratio of the similarity of the stimulus to exemplar in the category over the similarity of the stimulus to exemplars on all categories.  Thus, all objects are classified according to the degree of similarity they have to other objects in the category. This works for individual data, but cannot be integrated. Context theory uses the city block metric in which the similarity of two objects is measured by the sum of the absolute differences. This is true for separable (distinctive) features but it is much more difficult for integral features (such as luminescence and hue).

Context theory arises as a consequence of integrating the mapping hypothesis of the identification-classification relationship with models in the areas of choice and similarity. Context theory was related to a more general theoretical framework for the modeling of choice and similarity and the ratio-response rule for classification was related to Luce’s (1963) choice model for identification.

Bearing the different theories and models in mind, we attacked the question (which was to be raised many times during future discussions) of whether we have different memory systems or a single-system exemplar memory model. Exemplar theory assumes that categorization and recognition performance can be explained using a unified framework that depends on a single representation system. Although dissociations between categorization and recognition in amnesic and normal individuals has been interpreted as evidence of different memory systems, the argument was that this evidence can be interpreted as supporting a single-system exemplar memory model.  According to this argument, poor performance on recognition tasks but not on classification tasks for both amnesic patients and normal control subjects, who have had a delay between learning and testing, is due to memory sensitivity and not to multiple memory systems. That is, amnesic patients have lower memory sensitivity.

Questions to think about included: What new procedures could be made using these models? How would we use this model to predict better performance for amnesic patients?

Taking what we had learned about the theories and models of classification of objects and the different memory systems/single-system exemplar memory model debate we then examined how conceptual thought is grounded in perceptual similarity. Historically, there has been a division between conceptual and perceptual systems suggesting that perceptual similarities are not used when creating categories. If, however, conceptual processing shares computational resources with perception, than this suggests that a common representational and processing system underlies both. Two approaches were mentioned: the eliminative view and the agnostic view. The former advocated that human knowledge contains no non-perceptual representations and that amodal symbols have arbitrary relations to perception and to their referents in the world. The agnostic view differs from the eliminative view in its strictness.  It suggests that human knowledge has major perceptual components and may or may not contain non-perceptual components.  The agnostic view states that these amodal symbols may exist, but are not always necessary.  The agnostic view stresses the evolution of concepts from perception through processes that can eventually achieve abstract end states. There is strong evidence for this idea in studies that show that children rely on perceptual representations when developing their concept of numbers. 

 The discussion of amodal systems led to the discussion of language and its use as an abstract amodal system. This is especially relevant in reading as we at automatically process the symbols we perceive as information. Even in perceiving objects such as a coffee cups, we often do not define it by what it looks like, but rather by its function.

In the many roles of perception in conception, the primitive appeal of overall similarity became a point of discussion. For example, what is a bachelor and why do we have so much trouble describing one when there are simple rules we could use, but do not because they do not adequately describe the meaning of the concept? The appeal of overall similarity is in the fact that people make similarity judgments much more easily when objects are identical than when they have to decide whether they share a single property in common.

Another aspect of similarity studied this week dealt with how we process different stimuli. It is intuitive to assume that different types of processing are compatible with different types of stimuli.  Specifically, stimuli such as apples and oranges, which are taxonomically related (i.e., they belong to the same superordinate category) are compatible with comparison, while stimuli such as apples and baskets, which are thematically related (i.e., apples go in baskets), are more compatible with integration.  Stimuli that are taxonomically related are alignable (e.g., apple-orange).  Alignability refers to the ease of comparison, the judgment of similarities and differences.  Nonalignable stimuli (e.g., apple-basket, man-tie) may belong to categories that are based on thematic relations (e.g., events, places, scenes, etc.).

 Category formation is affected by both comparison and integration.  Models of induction also focus solely on feature comparison, yet there is clear evidence that thematic relations play a role.  There are also implications regarding metaphors, similes, and analogies and conceptual combination. The take-home message was clear. When categorizing concepts we need to include integration. Cognitive models must  include both integration and comparison processes and the effects of stimulus compatibility.  Even when clearly instructed otherwise, people enlist both integration and comparison processes depending upon the stimuli.  We are less likely to integrate concepts that are highly alignable. We compare by category first.  The kinds of stimuli can override task-related stimuli.