Cost of Attention as an Indicator of Category Learning

Cost of Attention as an Indicator of Category Learning Hyungwook Yim (yim.31@osu.edu) Department of Psychology & Center for Cognitive Science, The Ohio State University 209C Ohio Stadium East, 1961 Tuttle Park Place Columbus, OH 43210, USA Catherine A. Best (best.140@osu.edu) Department of Psychology & Center for Cognitive Science, The Ohio State University 208G Ohio Stadium East, 1961 Tuttle Park Place Columbus, OH 43210, USA Vladimir M. Sloutsky (sloutsky.1@osu.edu) Department of Psychology & Center for Cognitive Science, The Ohio State University 208C Ohio Stadium East, 1961 Tuttle Park Place Columbus, OH 43210, USA Rehder (2010) recorded eye movements during a supervised category learning task and found evidence for a cost of attention. If category learning involves selective attention, then a cost of attention could function as an indicator of category learning. Abstract Category learning often involves selective attention to category relevant information, which may result in learned inattention to category irrelevant information. This learned inattention is a cost of selective attention. In the current research, the cost of attention was used as an indicator of category learning. Participants were given a category learning task, and the amount of supervision given to them was manipulated. Along with behavioral data, recorded eye movements during the task showed signature patterns of learning via a cost of attention. In addition, a simple neural network (perceptron) was able to use these eye-tracking data to predict success in learning. Thus, the observed attentional pattern – the cost of selective attention – was proposed as an indicator of category learning. Overview of Current Experiments Keywords: category learning, cost of attention, eye tracking, supervised learning, sparse category, classifier, perceptron, neural network Introduction Attention plays a central role in many models of category learning (Kruschke, 1992; Nosofsky, 1986). During category learning, the ability to selectively attend to category-relevant cues while ignoring category-irrelevant cues allows for more efficient category learning. However, attention should also be flexible to enable learning of new categories. Consider learning to discriminate plums from nectarines. The most efficient way to distinguish them visually would be focusing on color as a cue rather than shape. However, when encountering a new category like lemons versus bananas, the once useful color cue no longer helps, while the previously unhelpful cue of shape becomes a good dimension to efficiently learn the categories. The process of ignoring the shape cue in the first learning instance often results in learned inattention to this cue (Kruschke & Blair, 2000). Learned inattention to a previously irrelevant dimension creates a deficit in future learning. This deficit constitutes a cost of attention (Hoffman & Rehder, 2010). For example, Hoffman & Two eye-tracking experiments were conducted with identical stimuli. Experiment 1 tested participants in a two- phase supervised category learning task that should promote learning. Experiment 2 tested participants in a two-phase unsupervised category learning task that should prevent learning. Critically, the second phase of each learning task relied on previously irrelevant cues to learn a new category. Based on previous research, it was predicted that supervision would facilitate category learning in Experiment 1, compared to unsupervised learning in Experiment 2. However, there is also evidence with adults that supervision only facilitates category learning when a category does not have much structure (Kloos & Sloutsky, 2008). Category structure could be measured as category density or “a ratio of variance relevant for category membership to the total variance across members and nonmembers of the category” (Sloutsky, 2010). Therefore, categories that have many features in common and those features are not shared with non-members are statistically dense. On the other hand, categories that have few features in common while having many features that are common with non-members are statistically sparse. Thus, in the current research an artificial sparse category was used to manipulate learning via supervision. The first aim of the current research was to replicate a condition with a cost of attention during category learning and a condition with no cost of attention during category learning by manipulating the amount of supervision provided to participants. The second aim was to use the demonstrated cost and lack of cost to classify adults into