Learning with a Purpose: The Influence of Goals

Learning with a Purpose: The Influence of Goals Sarah Wellen (swellen@andrew.cmu.edu) Department of Philosophy, Baker Hall 135 Pittsburgh, PA 15213 USA David Danks (ddanks@cmu.edu) Department of Philosophy, Baker Hall 135 Pittsburgh, PA 15213 USA Abstract Most learning models assume, either implicitly or explicitly, that the goal of learning is to acquire a complete and veridical representation of the world, but this view assumes away the possibility that pragmatic goals can play a central role in learning. We propose instead that people are relatively frugal learners, acquiring goal-relevant information while ignoring goal-irrelevant features of the environment. Experiment 1 provides evidence that learning is goal-dependent, and that people are relatively (but not absolutely) frugal when given a specific, practical goal. Experiment 2 investigates possible mechanisms underlying this effect, and finds evidence that people exhibit goal-driven attention allocation, but not goal- driven reasoning. We conclude by examining how frugality can be integrated into Bayesian models of learning. Keywords: Goals, Learning, Task-effects, Rationality, Frugality, Bayesian inference Introduction Intuitively, what we need to know depends on what we want to do: the information we require from the environment will partially depend on our goals, desires, and intentions. For example, consider reading a recipe. If I am deciding whether to make this dish for a friend with a dairy allergy, then I need to know simply whether the dish contains any milk at all. If I am instead preparing a shopping list so that I can later make the dish for myself, then I need to know how much milk is required, not just whether any at all is involved. In this paper, we examine the extent to which learning is responsive to pragmatic goals (e.g. our desire to succeed at an expected future task). Many cognitive models of learning assume that individuals are trying to acquire (approximately) complete representations of their environments, so pragmatic goals play essentially no role. For example, most models of causal learning assume that agents are trying to learn the “true” causal structure; most models of language learning assume that people are trying to infer the underlying structure of the language; and most models of category learning assume people are trying to acquire conceptual representations that most closely track the world’s statistical regularities. Under these models, pragmatic goals play essentially no role in learning; instead, the learner always tries to acquire a (relatively) complete and veridical representation of the world. This representation can later be used for a range of practical purposes precisely because it is complete and veridical. Despite this, previous research suggests that pragmatic goals do impact learning. For example, people acquire different categories from identical data when learning occurs through a categorization task (selecting a category label based on a set of feature values) vs. a feature inference task (inferring a feature value given the values of other features) (e.g., Markman & Ross, 2003; Zhu & Danks, 2007). Also, people learn more in dynamic control tasks when given a general learning goal (learn about the system) rather than a specific task (maintain the system at a specific state) (Burns & Vollmeyer, 2002; Osman & Heyes, 2005). Tasks have even been found to influence low-level processes; for instance, negative priming in selective attention is directed to only task-relevant dimensions of distractor objects (Frings & Wentura, 2006; Maruff et al., 1999; Tipper, Weaver, & Houghton, 1994). While task effects are common, there has been little study of the extent to which learning is modulated by longer-term pragmatic goals (vs. the task performed during learning). Much of everyday learning is driven by the desire to succeed at an expected future task, and it is possible that learning is highly responsive to beliefs about how information will be used in the future. If this is the case, our models of learning cannot ignore the important role that pragmatic goals play in many real-world learning situations. Our central theoretical proposal is that people’s pragmatic goals direct their learning towards pragmatically relevant information and, perhaps more importantly, away from pragmatically irrelevant information. That is, people are relatively frugal learners who encode only the information they need: they acquire goal-relevant representations and ignore goal-irrelevant dimensions of the environment. We first report experimental results suggesting that people are relatively frugal when given a concrete, pragmatic goal (Experiment 1). We then present preliminary evidence about possible mechanisms underlying this frugality (Experiment 2). We finish by arguing that frugality can be a ‘rational’ strategy that can be reconciled with commonly used models of rational learning, including Bayesian inference. Experiment 1 Experiment 1 directly tests whether people display frugal learning when provided with a concrete, practical goal. The learning paradigm involved four buttons that probabilistically produced numbers between 1 and 100, where two of the buttons had relatively high means and two