Categorization

I. Overview

A. Current State of Affairs

 Synergy in Cognitive Science: Categorization research is dynamic and exciting, resembling

cooperative algorithms in neural net models. Interactions across cognitive science subdisciplines
are increasing, addressing questions about rule application versus memory retrieval in
categorization.

 Interdisciplinary Motivations: Memory-based categorization in psychology aligns with case-

based reasoning in AI. Linguistics' belief in grammar rules is challenged by connectionist
researchers. Mutual interaction across disciplines is motivated by practical challenges.

B. Brief History

 Conceptual Structure: Central question involves the structure of concepts. Three positions:
classical view (defining properties), probabilistic view (graded membership), and exemplar view
(individual exemplars). Shift in the 1970s from classical to probabilistic view, notably influenced
by Eleanor Rosch's work on natural object categories.

 Models and Views: Different views of conceptual structure constrain category processing
models. Similarity-based models, criticized by some, explore constraints on the notion of
similarity. Current research focuses on contrasts and integration between similarity-based
learning (SBL) and explanation-based learning (EBL).

C. Ecological Validity and Artificial vs. Natural Categories

 Interplay and Tension: Psychological research on categorization involves both natural and
artificially constructed categories. Natural categories have lexical entries corresponding to real-
world entities, while artificial categories involve controlled experiments. Tension exists between
ecological validity and controlled experiments.

 Effective Interplay: Researchers use artificial categories to identify properties of natural

categories and vice versa. Examples include Rosch's research on typicality effects and Brooks'
studies on specific item similarity. Balancing ecological validity and controlled experiments is
crucial for meaningful results.

D. Summary

 Parallel, Interactive Research: Categorization research shows a parallel, interactive pattern. The
chapter organizes the review around various methods of inquiry, including evolutionary
considerations, philosophical perspectives, developmental and cross-cultural research,
computational models, and neuroscience.

 Overview of Methodologies: The chapter aims to provide a comprehensive overview, exploring

various methodologies to convey the richness and strengths of categorization research.

 Future Discussion: The last section will summarize and integrate findings, discussing challenges
and opportunities in categorization research.
II. Methods of Inquiry

A. Evolutionary

1. Conceptual Functions:

 Categorization: Determines category membership.

 Understanding: Relates new information to existing knowledge.

 Learning: Supports updating of knowledge.

 Inference(推理): Allows predictions about behavior.

 Explanation: Supports reasoning and persuasion.

 Conceptual Combination: Forms new concepts from existing ones.

 Planning: Instantiates goals in the planning process.

 Communication: Utilizes concepts for effective communication.

2. Purpose of Categorization:

 Anderson's Rational Model:

 Assumes the mind is a rational, optimal system.

 Optimizes predictability of features for categorization.

 Analyzes environmental information structure.

 Derives an iterative algorithm for optimal categorization.

 Successful in predicting artificial category learning experiments.

 Critique:

 Challenges assumptions about information structure.

 Highlights the importance of real-world learning and use.

 Shepard's Universal Law of Generalization:

 Focuses on the generalization task faced by all organisms.

 Derives a mathematical account of generalization.

 Proposes a negative exponential function for generalization probability.

 Supported by studies showing a similar functional form.

3. Summary:

 The evolutionary perspective considers the adaptive nature of cognition.

 Anderson's rational model offers insights but faces critiques regarding assumptions.
  Shepard's work emphasizes the universal task of generalization and aligns with
descriptive psychological models.

 William James' insight on the mutual evolution of mind and world underscores the
ongoing debate in categorization research.

 Addressing conceptual functions remains a key principle in categorization research.

Philosophical Perspective:

1. Reference:

 The text discusses Putnam's argument against the traditional view that concepts mediate the
link between words and referents.

 Putnam suggests that direct reference, without a conceptual representation, may be possible,
especially for certain types of nouns like natural kind terms.

 Concepts may differ in the beliefs about the completeness and validity of properties, impacting
reference.

2. Kinds of Categories:

 Putnam distinguishes between natural kind terms (discovered by science), nominal kind terms
(established by convention), and artifact categories.

 Categories may face tension when entities exhibit surprising properties, leading to a conflict
between modifying concepts and beliefs.

3. Meaning:

 The distinction between metaphysics (concerned with how the world is) and epistemology
(concerned with how we know the world) is emphasized.

 The study of concepts focuses on epistemological questions, leaving the determination of

reference for natural kinds to experts.

4. Summary:

 The discussion highlights the centrality of questions about relationships among category
representations, reference, and meaning in understanding categorization.

Developmental Perspective:

1. Similarity and Development:

 Developmental changes in similarity processing are linked to categorization.

 Young children may process stimuli holistically, while older children are more analytic.

2. Constraints, Theories, and Development:

  Rejecting the "doctrine of original sin," some argue that even young children's category learning
is constrained by domain-specific theoretical biases.

 Evidence supports the idea that children reason in an essentialist manner, assuming hidden
causal mechanisms.

3. Category Labels and Learning:

 Children place constraints on reference, assuming novel terms apply to entire objects, expecting
objects to have only one label, and assuming labels refer to objects of like kind.

4. Kindhood and the Basic Level:

 The concept of the basic level is introduced, indicating the preferred, most abstract, and rapidly
categorized level.

 Notions of kindhood may correspond to the basic level, but there's debate on whether it
changes with development or expertise.

5. Domain Specificity and Cross-Domain Interactions:

 Psychological essentialism is mainly examined for biological kinds, but there's discussion about
whether similar beliefs exist for artifacts, personality traits, etc.

 Questions arise about cross-domain interactions and whether children generalize beliefs about
underlying natures from biological to other kinds.

6. Summary:

 The text underscores the significance of developmental research in understanding

categorization, particularly in integrating knowledge (theories) and experience in learning.

Cross-Cultural Comparisons in Categorization Research:

 Approach-avoidance conflict for researchers.

 Logistically challenging and potentially confounded.

 Provide a powerful test for the generality of observations or principles.

 Limited use in categorization research but impactful.

1. Basic Levels:

 Rosch's basic level analyses influenced by cross-cultural comparisons.

 Berlin, Breedlove, and Raven found cross-cultural agreement at the folk generic level.

 Folk generic level corresponds to the genus level of scientific taxonomy.

 Ethnobiologists suggest consistency due to correlated features in organisms.

  Disagreement between ethnobiological and psychological measures of basic level.

 Puzzle: Berkeley undergraduates vs. anthropological investigations.

 Basic level may change with expertise or involve divergent criteria.

2. Similarity-Based vs. Theory-Driven Learning:

 Cross-cultural ideal testing grounds for SBL vs. EBL.

 SBL models sensitive to types and token frequency.

 Differences may result from familiarity or differential experience.

 Atran argues for a cross-cultural universal essentialist stance toward biological kinds.

3. Cultural Consensus Model (CCM):

 Tests agreement among informants based on cultural consensus.

 Assumes a single consensus; informants' answers are independent.

 Competence parameter reflects the probability of knowing the consensus.

 Useful tool for evaluating agreement and disagreement across cultures.

4. Anthropological Studies on Biological Kinds:

 Focus on biological kinds shows cross-cultural consistency.

 Substantial agreement of folk biology with scientific taxonomy.

 Expertise associated with deviations from scientific taxonomy.

5. Use of Categories in Reasoning:

 Limited research on categories in reasoning across cultures.

 Walker's study in Nigeria on preservation-of-identity judgments.

 Nonpreservation judgments influenced by centrality to ritual practices.

6. Optimistic Outlook on Cross-Cultural Comparisons:

 Strong cross-cultural similarities in categorization.

 CCM and category-based induction model are valuable methodological tools.

 Cross-cultural comparisons not a panacea but a fruitful avenue for exploration.

E. Computational Models in Categorization Research:

1. Issues and Purposes of Modeling:

 Tradition of implementing theoretical ideas as computational or mathematical models.

  Categorization models serve as precise statements of categorization accounts.

 Models act as a common language to compare different theories of categorization.

 Modeling helps uncover potential errors in human reasoning.

 Models offer insights that may not be apparent with verbal descriptions.

 Example: Nosofsky demonstrated equivalence between Anderson's rational theory and Medin's
context theory.

 Computational models, so far, are not complete accounts of categorization.

2. Limitations and Virtues of Modeling:

 No computational model has provided a comprehensive account of categorization.

 Models often focus on critical aspects of theory, leaving other issues in the background.

 Example: Busemeyer et al. showed that certain connectionist models fail to account for the cue
competition effect in human category learning.

 Models allow focusing on critical aspects of theory while ignoring non-essential details.

3. Integration with Other Theoretical Accounts:

 Categorization models should not be isolated but integrated with other cognitive theories.

 Categorization is interconnected with learning, memory, and reasoning.

 Synergy between categorization models and other computational models, e.g., exemplar models
and multiple-trace models of memory.

 Connectionist modeling offers another framework for developing general models of

categorization and cognitive abilities.

E. Computational: Issues and Purposes of Modeling

 Purpose of Modeling:

 Models in categorization serve as precise statements of theoretical ideas.

 Acts as a common language for describing theories, facilitating comparisons.

 Allows for the implementation and comparison of different accounts of categorization.

 Provides a way to address potential errors in human reasoning.

 Insurance Against Errors:

 Computational modeling acts as insurance against errors in predicting theory outcomes.

 Difficult for researchers to predict theoretical predictions until implemented in a model.

 Incomplete Models:
  No computational model has provided a complete account of categorization.

 Models often assume representation of category members using features but lack
explanations for learning these features.

 Synergy Between Models:

 Categorization models should ideally integrate with other theoretical accounts of

cognition.

 Examples of synergy include connections between exemplar models of categorization

and multiple-trace models of memory.

Learning

 Categories Overview:

 Models generally fall into two categories: Similarity-Based Learning (SBL) and
Explanation-Based Learning (EBL).

 Similarity-Based Learning (SBL):

 Abstraction Models:

 Represent concepts with summary information about category members.

 Historical importance, but limited as people learn more than central tendencies.

 General Recognition Theory represents perceptual categories with probabilistic

density functions.

 Exemplar Models:

 Represent concepts with memories of specific category members.

 Context model predicts learning of non-linearly separable categories.

 Generalized Context Model (GCM) improves representation flexibility.

 Applications in AI through case-based reasoning.

 Connectionist Models:

 ALCOVE model combines aspects of exemplar and connectionist models.

 Focus on attentional weights on different dimensions.

 Gluck and Bower's model and its extension (configural cue model).

 Rule-Based Models:

 Rule-based approaches had attention in the past.

 Michalski's INDUCE system induces rules from partitioned examples.

  Discrimination net models involve branching tests.

 Rules often over-generalize initially and get refined.

 Comparing Models:

 Similarity-based models are often compared within frameworks rather than

across frameworks.

 Different models within the same framework may share more similarities than
models in different frameworks.

 Explanation-Based Learning (EBL):

 Addresses the limitations of bottom-up, data-driven processes in similarity-based

models.

 People bring prior knowledge and theories to influence concept formation.

 AI models of EBL use background knowledge to explain and generalize from training
examples.

 Recent applications of EBL to psychological experiments.

 Integration of SBL and EBL:

 Growing interest in methods for integrating similarity-based learning and explanation-

based learning.

Merging SBL and EBL: Integration Model

 Integration Model Overview:

 Modification of a similarity-based model to incorporate the effects of prior knowledge.

 Integration model is an exemplar model, a variant of context theory (Medin & Schaffer,
1978).

 Influence of Prior Knowledge:

 Two kinds of exemplars influence categorization judgments: category-specific exemplars

and prior examples from other categories.

 Example scenario: Learning about joggers in a new city influenced by prior examples of
joggers from previous contexts.

 Over time, categorization judgments shift from being influenced by prior examples to
being influenced more by observed examples in the new context.

 Comparison to Bayesian Models:

 Similar to Bayesian models of statistical estimation, considering the changing influence

of prior knowledge over time.
 Additional Processes by Which Prior Knowledge Affects Learning:

1. Selective Weighting of Category Members:

 Observations that fit prior knowledge are remembered better.

 Example: Better learning about joggers with expensive running shoes.

2. Selective Weighting of Features:

 Prior knowledge leads to attention on specific features.

 Example: Attending to people's shoes instead of hair color in categorizing

joggers.

3. Distortion Effect:

 Prior knowledge may distort perceptions.

 Example: A jogger without expensive running shoes may be misremembered as

having them.

 Research Findings:

 Heit (1994) found that the integration model provided a good qualitative and
quantitative account in experiments simulating category learning with prior knowledge.

 Other processes (selective weighting, distortion effect) were plausible but not necessary
to explain results.

 Other Approaches to Integration:

 UNIMEM System (Lebowitz):

 SBL determines regularities, and UNIMEM explains these commonalities with its
domain theory.

 IOU, IOE, EXOR Systems:

 IOU and IOE operate with EBL followed by SBL or vice versa.

 EXOR combines EBL and SBL for predictive features and explanations.

 Interaction Between EBL and SBL:

 Unidirectional and indirect interaction common in integrated systems.

 Some argue for more tightly coupled systems, especially in psychological process
models.

 Wisniewski and Medin's Argument:

 Advocacy for greater interaction between theory and data.

  Modular ways of incorporating prior knowledge into categorization models may be
inadequate.

 Observations on features, rules, and learning history variability support the need for
more interactive models.

3. Induction: Category-Based Induction (CBI) Model

 CBI Model Overview:

 Focuses on the inference and inductive reasoning function of concepts.

 Addresses how people infer novel properties of categories.

 Factors Influencing Inductive Soundness:

1. Similarity:

 Critical for evaluating inductive soundness.

 People project new properties based on the similarity between premise and
conclusion categories.

 Example: Likelihood of dogs having sesamoid bones based on the premise about
cows.

2. Coverage of the Premise:

 Similarity between premise category and members of the superordinate

category.

 Example: Inductive inference from cow to dog based on the category "mammal."

 Role of Categories in Inductive Reasoning:

 CBI model combines similarity and coverage to evaluate inferences.

 Atypical category members contribute less to coverage.

 Property-Related Considerations:

 CBI model uses "blank properties" in studies (e.g., has sesamoid bones).

 Heit and Rubinstein (1994) propose that the property being inferred influences attention
to relevant properties when evaluating similarity.

4. Conceptual Combination: Selective Modification Model

 Selective Modification Model Overview:

 Addresses conceptual combination, specifically adjective-noun combinations.

 Example: Understanding "brown apple" involves modifying the prototype of "apple"

with the dimension of color.
  Complexities of Conceptual Combination:

 People use general knowledge about feature relations.

 Conceptual combination can lead to inferences about emergent features not present in
constituent concepts.

 Example: Expecting a nonconformist trait in a Harvard-educated carpenter.

 Compositional Process:

 Conceptual combination may be a compositional process.

 Additional knowledge beyond constituent concepts is needed for interpretation.

5. Summary: Challenges and Opportunities

 Challenges:

1. Questions of Features:

 Little progress in understanding how features come into existence.

 Need to explore feature construction in conceptual properties.

2. Question of Kinds:

 Narrow focus on certain stimulus materials and tasks raises challenges of

generality and misattribution.

 Limited exploration of categories beyond nominal, natural kind, and artifact

categories.

3. Questions of Structure:

 Limited attention to relational properties and their role in creating structure.

 Need to consider structural relations between features and category members.

4. Questions of Learning:

 Importance of cognitive development and conceptual change in categorization

research.

 Lack of linguistic perspective in the review.

 Opportunities:

 Integration of perspectives from linguistics, computational vision, and other areas for
more ambitious categorization research.