Effect of memory decay on predictions from changing categories.

Abstract

In contrast to the static categories assumed in most categorization experiments, many real-world categories undergo gradual and systematic change in their definitions over time. Four experiments were carried out to study such category change. In these studies, participants successfully adjusted as category change occurred, but also showed a lingering and cumulative effect of past observations. The participants' performance was closely modeled by incorporating memory decay for past observations into J. R. Anderson's (1990, 1991) rational categorization algorithm and into a version of R. M. Nosofsky's (1986) exemplar categorization model. The resulting models suggest that the decay function is closer to a power law than to an exponential and that decay occurs both by item and by time, with the item decay being stronger than the time decay. The finding of power law decay gives additional support to claims that exemplar memories are used in categorization.

Learning bundles of stimuli renders stimulus order as a cue, not a confound

“…The effect of stimulus order on cognition has often been viewed as a consequence of the way that the human mind processes information (1)(2)(3)(4)(5)(6)(7)(8). For example, Elliott and Anderson (5) accounted for the effect of stimulus order on category learning by incorporating the mechanisms of memory decay into a categorization model.…”

“…For example, Elliott and Anderson (5) accounted for the effect of stimulus order on category learning by incorporating the mechanisms of memory decay into a categorization model. Memory decay intuitively captures the assumption that the human mind tends to be influenced by nearby stimuli more than by distant ones.…”

Learning time-varying categories

“…However, the changeable nature of many of the concepts and categories with which humans must interact has not been greatly emphasized in the categorization literature (but see Elliott & Anderson, 1995). In category-learning experiments, it is generally assumed that the underlying category is more or less static, and as such, the order in which one encounters category members should not matter to a rational learner.…”

“…Many studies have avoided any explicit discussion of whether this sensitivity should be called normative (e.g., Stewart et al, 2002), others have argued that it reflects the cognitive limitations of the human learner (Sakamoto et al, 2008;Sanborn et al, 2010), and still others have suggested that order-sensitive learning rules are necessary if the learner is to be able to adapt to a changing world (Elliott & Anderson, 1995;Nosofsky et al, 1992). The latter perspective is mirrored rather explicitly in the literature on sequential effects (Yu & Cohen, 2009) and change detection (Brown & Steyvers, 2009).…”

“…In most models, the weight assigned to a particular observation tends to decay approximately exponentially as a function of age (Brown & Steyvers, 2009;Nosofsky et al, 1992;Yu & Cohen, 2009), although in some cases the decay has been proposed to be a power function (Elliott & Anderson, 1995), which would be more in keeping with the literature on memory (Rubin, Hinton, & Wenzel, 1999;Wixted & Ebbesen, 1997). A similar outcome exists in the judgmental forecasting literature, which examines how people perceive and extrapolate time series data: Once again, an exponential weighting rule appears to account for human performance (see Wright, 1994, andLawrence, Goodwin, O'Connor, &Onkal, 2006, for overviews).…”