Category Learning Through Active Sampling

Category Learning Through Active Sampling Doug Markant (doug.markant@nyu.edu) Todd M. Gureckis (todd.gureckis@nyu.edu) New York University, Department of Psychology 6 Washington Place, New York, NY 10003 USA Abstract ence relates to different hypotheses (c.f., Bruner, 1961). In a study of active intervention during a causal learning task, Sobel and Kushnir (2006) showed that active learners were more likely to learn a hidden causal structure than participants that were “yoked” to their interventions (i.e., a group with the same data but who did not independently make sampling decisions). Similar concerns are often used to support edu- cational practices that emphasize “inquiry” or “discovery”- based instruction (Kuhn et al., 2000). The aims of the present study were two-fold. First, we were interested if participants could adaptively structure their own learning experiences when acquiring new concepts. Second, we were interested in how the effectiveness of active sam- pling might interact with the specific structure of categories. While a number of recent studies have explored how learners make information sampling decisions to support their own learning (Castro et al., 2008; Kruschke, 2008; Gureckis & Markant, 2009; Steyvers et al., 2003), there has not yet been a systematic evaluation of how this ability might vary across different category structures. Laboratory studies of human category learning tend to empha- size passive learning by limiting participants’ control over the information they experience on every trial. In contrast, we explore the impact that active data selection has on category learning. In our experiment, participants attempted to learn categories under either entirely passive conditions, or by ac- tively selecting and querying the labels associated with par- ticular stimuli. We found that participants generally acquired categories faster in the active learning condition. Furthermore, this advantage depended on learners actually making decisions about which stimuli to query themselves. However, the effec- tiveness of active sampling was modulated by the particular structure of the target category. A probabilistic rule-learning model is proposed that explains the results in terms of a strong prior bias towards uni-dimensional rules which impairs learn- ing of alternative category boundaries. Active learners appear to be able to bootstrap their own learning, but this ability may be strongly constrained by the space of hypotheses that are un- der consideration. Keywords: categorization, active learning, information sampling, rule learning, decision-bound models Despite the widely held view that people learn better by do- ing than simply observing, there have been surprisingly few detailed accounts of the impact that “active” information ac- quisition has on the learning process. In particular, theoret- ical models which explain how people learn new concepts from examples usually treat learners as passive accumulators of evidence about the structure of categories. For example, the standard procedure in most category learning experiments is to exhaustively and randomly sample the set of training stimuli. However, in everyday life, human learners can often control their own learning by selectively “sampling” partic- ular observations they estimate to be useful or informative. The goal of the present paper is to understand the cognitive consequences of this type of learning. There are at least two explanations for why active sampling might result in better learning than passive observation. First, rather than being limited by the flow of information from pas- sive experience, active learners are free to select which infor- mation they want to learn about. For example, by making directed queries that take into account their current uncer- tainty, the learner may be able to optimize their experience (e.g., avoiding redundant data). Research in machine learn- ing has shown that the principle of uncertainty sampling (se- lectively querying data that is expected to be informative) can have a dramatic impact on the amount of training needed to reach a performance criterion (Settles, 2009). Independent of the advantage of better data, active learn- ers may also benefit from greater engagement in the learning task. For example, the very act of planning interventions or deciding which samples to take may necessitate deeper eval- uation of the problem structure and of how observed experi- Overview of the present experiment Our experiment adapts a well-studied paradigm for percep- tual category learning using multidimensional, continuous- valued stimuli. In the task, participants learned to classify perceptual stimuli into different abstract groups. Two types of category structures were used: rule-based (RB), in which the decision rule is defined as a criterion along a single di- mension, and (2) information-integration (II), in which the decision rule is a function of at least two dimensions (see Figure 1). Participants in the experiment were further divided into three training conditions. In the passive-normal condi- tion, participants observed training stimuli that were gener- ated from two bivariate normal distributions (i.e., a standard training procedure). In the active condition, participants were able to “design” stimuli for which they received feedback about the category label. In the passive-yoked condition, each participant was linked to an active learner, passively observ- ing the samples they made and receiving the same feedback. There are three key aspects of the design worth highlight- ing. First, in binary classification tasks, the optimal sampling strategy is simply to make queries close to the current es- timate of the category boundary (or margin) — the region of greatest uncertainty. However, we anticipated that partic- ipants’ ability to do so might vary between the RB and II learning tasks. Previous research has suggested that these two types of category structures may be learned in fundamentally different ways (Ashby, Alfonso-Reese, Turken, & Waldron,