Human Categorization Learning as Inspiration
     for Machine Learning Algorithms

                         Christina Zeller and Ute Schmid

                  Cognitive Systems Group, University of Bamberg
                    An der Weberei 5, 96045 Bamberg, Germany
                 {Christina.Zeller,Ute.Schmid}@uni-bamberg.de

    Empirical observations of humans learning to categorize inspired the devel-
opment of early machine learning algorithms (cf. Unger & Wysotzki, 1981). For
example, Hunt, Marin, and Stone (1966) developed a decision-tree learning algo-
rithm based on experiments by Bruner, Goodnow, and Austin (1956). However,
nowadays the focus of machine learning lies on efficient categorization and not
on cognitive plausibility of the underlying learning algorithms.
    Recently Lafond, Lacouture, and Cohen (2009) modeled human categoriza-
tion behavior with decision-trees, but they did not address the question of how
these decision-trees are constructed from training trials. We analyzed their data
and could show that a measure of incremental information gain can be an ap-
propriate feature selection criterion (Zeller & Schmid, accepted).
    Empirical data imply that humans use (meta-)strategies while learning to
categorize (cf. Unger & Wysotzki, 1981). Typically, humans focus first on single
features as categorization criteria and only later use conjunctions or disjunctions.
As a next step, we plan to conduct case studies where participants, while learning
correct categorization with a trial by trial feedback, shall explain their decisions.
Based on the results, we intend to design experiments where the material is
constructed in such a way that cues enhance or hinder specific strategies. We
hope that a deeper understanding of the human process of categorization learning
can inspire cognitively plausible machine learning algorithms.

References
Bruner, J. S., Goodnow, J. J., & Austin, G. A. (1956). A study of thinking: With
      an appendix on language by Roger W. Brown. New York, NY: Wiley.
Hunt, E. B., Marin, J., & Stone, P. J. (1966). Experiments in induction. New
      York, NY: Academic Press.
Lafond, D., Lacouture, Y., & Cohen, A. L. (2009). Decision-tree models of cat-
      egorization response times, choice proportions, and typicality judgments.
      Psychological Review , 116 , 833–855.
Unger, S., & Wysotzki, F. (1981). Lernfähige Klassifizierungssyssteme (Classifi-
      cation Systems Being Able to Learn). Berlin, Germany: Akademie-Verlag.
Zeller, C., & Schmid, U. (accepted). Rule learning from incremental presentation
      of training examples: Reanalysis of a categorization experiment. In 13th
      Biannual Conference of the German Cognitive Science Society (Bremen,
      Germany, Sept. 2016).