Academic journal article Psychonomic Bulletin & Review

Revealing Human Inductive Biases for Category Learning by Simulating Cultural Transmission

Academic journal article Psychonomic Bulletin & Review

Revealing Human Inductive Biases for Category Learning by Simulating Cultural Transmission

Article excerpt

Published online: 7 January 2014

© Psychonomic Society, Inc. 2014

Abstract We explored people's inductive biases in category learning-that is, the factors that make learning category structures easy or hard-using iterated learning. This method uses the responses of one participant to train the next, simulating cultural transmission and converging on category structures that people find easy to learn. We applied this method to four different stimulus sets, varying in the identifiability of their underlying dimensions. The results of iterated learning provide an unusually clear picture of people's inductive biases. The category structures that emerge often correspond to a linear boundary on a single dimension, when such a dimension can be identified. However, other kinds of category structures also appear, depending on the nature of the stimuli. The results from this single experiment are consistent with previous empirical findings that were gleaned from decades of research into human category learning.

Keywords Category learning . Bayesian modeling . Mathematical models

The ability to learn new categories from labeled examples is a basic component of human cognition, and one of the earliest to be studied by psychologists (Hull, 1920). As with many cognitive tasks, category learning can be characterized as a form of induction-an inference to underdetermined hypotheses from limited data (Bruner, Goodnow, & Austin, 1956). This directly implies a role for what machine learning researchers refer to as the induc- tive biases of a learner-those factors that make a learner more likely to entertain one hypothesis than another (Mitchell, 1997). In category learning, these inductive biases determine whether a particular category structure is easy or hard to learn.

Inductive biases for category learning have often been studied using "supervised" learning tasks, in which partici- pants learn experimenter-designed categories in order to study the rates at which they learn and how they generalize (e.g., Ashby, Alfonso-Reese, Turken, & Waldron, 1998; Erickson & Kruschke, 1998;Nosofsky,1986, 1987; Nosofsky, Palmeri, & McKinley, 1994). However, this is an inefficient way to study human inductive biases. Establishing what is easy or hard to learn requires teaching people many different kinds of categories and comparing the results (e.g., Shepard, Hovland, & Jenkins, 1961). Over the last few decades, a clear picture has emerged from pursuing this meth- od, but it has taken many experiments to produce this picture.

An alternative approach has been to use an "unsupervised" learning task, in which participants are asked to organize a set of stimuli into categories by themselves in order to examine which structures they naturally identify (e.g., Ahn & Medin, 1992; Anderson, 1991;Handel&Imai,1972;Imai&Garner,1965; Love, Medin, & Gureckis, 2004; Medin, Wattenmaker, & Hampson, 1987; Milton & Wills, 2004; Pothos & Chater, 2002; Pothos et al., 2011; Regehr & Brooks, 1995). Unsuper- vised learning clearly reflects people's inductive biases, but arguably taps different cognitive processes than does supervised learning (Milton & Wills, 2004; Regehr & Brooks, 1995).

For this article, we used a new approach to investigate human category learning, in which participants performed a standard supervised learning task, but we manipulated the categories that they learned so as to directly reveal their inductive biases using a single, compact experiment. The novel method that we used is known as iterated learning and has its roots in accounts of language evolution (Kirby, 2001). Participants are arranged into a chain in which the responses from the first participant are used as training data for the second participant, and so on. The result is a simple simulation of the cultural transmission of information that can be conducted in the laboratory. Mathematical analyses of this process have shown that as the chain gets longer, the re- sponses that people produce come to reflect their inductive biases (Griffiths & Kalish, 2007). …

Search by... Author
Show... All Results Primary Sources Peer-reviewed

Oops!

An unknown error has occurred. Please click the button below to reload the page. If the problem persists, please try again in a little while.