Working Memory Capacity and Generalization in Predictive Learning

Working Memory Capacity and Generalization in Predictive Learning Andy J. Wills (a.j.wills@ex.ac.uk) Psychology, University of Exeter, Perry Road, Exeter EX4 4QG. UK Thomas J. Barrasin (tombarrasin@hotmail.com) Psychology, University of Exeter, Perry Road, Exeter EX4 4QG. UK Ian P. L. McLaren (I.P.L.McLaren@ex.ac.uk) Psychology, University of Exeter, Perry Road, Exeter EX4 4QG. UK Abstract The relationship between working memory and deliberative processing was examined in a human contingency learning experiment that employed the combined positive and negative patterning procedure of Shanks and Darby (1998). Participants with a large working memory capacity showed generalization consistent with the application of an opposites rule (i.e., a compound and its elements signal opposite outcomes), whilst individuals with a small working memory capacity showed generalization consistent with surface similarity. Working memory capacity was assessed via the Operation Span task (Turner & Engle, 1989). Implications for associative, inferential, and dual-process accounts of human learning are discussed. Keywords: rules; associative learning; working memory; deliberative processing. generalization; Introduction The distinction between deliberative and non-deliberative processing, under a variety of different names, is fundamental to the study of cognition. For example, theorists seek to distinguish between propositional and associative learning (Mitchell, De Houwer & Lovibond, 2009), between analytic and nonanalytic categorization (Brooks, 1978), between automatic and intentional retrieval from memory (Jacoby, 1991), and between intuitive and deliberate reasoning (Kahneman, 2003). Deliberative processing is generally considered to be characteristic of thought processes that go beyond surface similarity to extract casual (De Houwer & Beckers, 2003) or abstract (Shanks & Darby, 1998) structure, thought processes that go beyond simple familiarity to episodic recollection (Jacoby, 1991), thought processes that are able to detect and correct irrational non-deliberative inferences (Kahneman, 2003). Deliberative thought processes are also often considered to be those that involve a degree of recurrence – in the sense that one goes through a series of intermediate stages to arrive at the final response (Milton & Wills, 2004). Another, related, way of capturing this idea of recurrence is to say that deliberative thought approximates the operation of a physical symbol system (Newell, 1980) – the ideas are related because certain recurrent, neural-like, structures have been shown to be able to implement a Universal Turing Machine (Siegelmann & Sontag, 1995). In the current study we investigated the relationship between the availability of working memory resources and the extent to which people engage in deliberative processing when acquiring new information. People with comparatively large working memories learn some tasks more quickly (e.g. learning to trace electrical signals through logic gates; Kyllonen & Stephens, 1990), and other tasks more slowly (e.g. acquisition of a hard-to-verbalize category structure; De Caro, Thomas & Beilock, 2008; but see Tharp & Pickering, 2009), than people with comparatively small working memories. In the current study, we were interested primarily in the relationship between availability of working memory resources and the nature of what was learned. One related study is that by De Houwer and Beckers (2003). In their forward cue competition experiment, participants first observed (in a computer game scenario) that firing a particular weapon (A) was followed by the destruction of a tank. Later on, weapon A was fired simultaneously with a new weapon, B. This compound firing also led to the destruction of the tank. They were then asked about the causal status of weapon B with respect to the destruction of the tank. On one, non-deliberative, account weapon B and the destruction of the tank have been repeatedly paired and thus one might say weapon B causes destruction of the tank by the mere fact of contiguity. However on another, deliberative, account one might argue that the causal status of B is uncertain because A causes destruction of the tank on its own, and B has never been used on its own. De Houwer and Beckers found that the imposition of a concurrent working memory load led to higher ratings for the extent to which B was considered to cause the tank’s destruction, compared to a situation where the same contingencies were observed in the absence of a concurrent load. Their conclusion was that the imposition of a concurrent load interfered with the deliberative (deductive reasoning) processes required to work out that the causal status of B was uncertain, despite the fact it had been repeatedly paired with the destruction of the tank.