Propositions, memories, and their interaction
Mitchell et al. use the simple example of learning that a bell signals food to illustrate the differences between the propositional and dual-system approaches. Although useful, the simplicity of the example is potentially deceptive. Focusing on a situation in which the organism learns about a single binary cue (bell rings/does not ring) and a single binary outcome (food presented/not presented) potentially leaves out some of the devilish details of how organisms learn in multiple-cue environments. Learning to form the proposition: “When the bell rings, I expect food” does not appear too arduous (for humans and some other species at least); but is the same kind of propositional statement the only form of knowledge learned when situations become more complex?
Imagine an environment where cues and outcomes are continuous and relations are probabilistic. In such an environment, do organisms form propositions of the kind: “When the bell rings for more than 5 seconds (but not over 15 seconds), the green light is at 50% brightness, and the red light is off, I expect food approximately 80% of the time”? Research into multiple-cue-probability learning (Enkvist et al. Reference Enkvist, Newell, Juslin and Olsson2006; Juslin et al. Reference Juslin, Olsson and Olsson2003), multi-attribute judgment (Newell & Bröder Reference Newell and Bröder2008), and categorization and concept learning (e.g., Allen & Brooks Reference Allen and Brooks1991; Nosofsky et al. Reference Nosofsky, Clark and Shin1989) has suggested that humans might try to learn such propositional information (i.e., rules) up to a point, but if the environment is too complex (e.g., cue-outcome relations are nonlinear), or feedback is insufficient or inappropriate, other forms of knowledge – principally stored instances – are relied upon.
Mitchell et al. acknowledge that instance memories play a role (sect. 3.1) but state that “recollections of past bell-food pairings alone cannot produce learning” (sect. 3.1, para.6). Such a conclusion implies that experiments demonstrating behaviour accounted for by an exemplar model (which relies exclusively on stored representations of stimuli; e.g., Juslin et al. Reference Juslin, Olsson and Olsson2003; Nosofsky et al. Reference Nosofsky, Clark and Shin1989) are not demonstrations of learning. This conclusion seems too extreme. Participants in such experiments have learned to classify particular objects as belonging to Category A and others to Category B – they have learned an association between a stimulus (the to-be-classified-object) and a response (the category label). But the content of this learning appears to be instances rather than a proposition (see also Shanks & St. John Reference Shanks and St. John1994). The interplay (and relative influence) of instances and propositions is somewhat underspecified in Mitchell et al's approach. However, the implication is that learning can only occur when propositions (rules) are formed. This seems a step too far, especially in situations with multiple, non-binary cues.
Systems, processes, and their interaction
Mitchell et al. note that many dual-system models do not specify how systems interact with each other. This is certainly true and, moreover, it appears that there is little consensus across different areas on how such interaction might occur. For example, an influential dual-system model of category learning, COVIS (Ashby et al. Reference Ashby, Alfonso-Reese, Turken and Waldron1998), proposes an initial bias towards an explicit hypothesis testing system, which is then usurped by an automatic, procedural system when the explicit system fails to learn. In contrast, popular dual-system theories of reasoning (e.g., Evans Reference Evans2008) suggest that the initial bias is towards the automatic, intuitive system, which is only corrected by the explicit system when things appear to go awry. Part of the problem in specifying these interactions is that it is often not clear what is meant by a system or a process – and whether these terms are interchangeable (Evans Reference Evans2008).
Mitchell et al. make a distinction (sect. 3.1), stating that their propositional approach is not a dual-system approach but that there are two types of processes (automatic processes of perception and memory and non-automatic processes of reasoning) in their single learning system. Although such specification clearly distinguishes their approach from the link-mechanism theories they wish to challenge, it blurs the distinction with many of the dual-process approaches to “higher-order” cognition. In their footnote, Mitchell et al. contrast their approach to other “dual-process or dual-system” theories of reasoning by stating that such approaches focus on performance, not learning – but at the same time Mitchell et al. want to incorporate memory (i.e., performance) processes into their learning system. Perhaps some clarification could be achieved by defining exactly what Mitchell et al. mean by a system (cf. Sherry & Schacter Reference Sherry and Schacter1987).
New techniques for old problems
Much of the evidence for and against the propositional and link approaches reviewed in the target article comes in the form of task dissociations; that is, situations in which a variable (e.g., cognitive load, instruction) is claimed to have an effect on one system but no effect on another. However, dissociations are unable to bear the inferential weight placed upon them for several reasons. One reason is that a simple dissociation requires that a variable have no effect on a particular behavioural measure, an assertion that is impossible in principle to verify (Dunn Reference Dunn2003). Hence, although dissociations may be found, they are neither necessary nor sufficient for drawing inferences about the number of processes or systems underlying observed behaviour (Newell & Dunn Reference Newell and Dunn2008). An alternative technique which avoids the flaws of dissociation logic is state-trace analysis (Bamber Reference Bamber1979). This technique has already been applied successfully to many areas of cognitive science (see Newell & Dunn Reference Newell and Dunn2008), and its application to the areas reviewed in the target article might prove fruitful.
