Associations and propositions: The case for a dual-process account of learning in humans

Abstract

N.J. Mackintosh (2104). Associations and Propositions: The case for a dual-process account of learning in humans. Neurobiology of Learning and Memory, 108,. AbstractWe review evidence that supports the conclusion that people can and do learn in two distinct ways -one associative, the other propositional. No one disputes that we solve problems by testing hypotheses and inducing underlying rules, so the issue amounts to deciding whether there is evidence that we (and other animals) also rely on a simpler, associative system, that detects the frequency of occurrence of different events in our environment and the contingencies between them. There is neuroscientific evidence that associative learning occurs in at least some animals (e.g., Aplysia californica), so it must be the case that associative learning has evolved. Since both associative and propositional theories can in principle account for many instances of successful learning, the problem is then to show that there are at least some cases where the two classes of theory predict different outcomes. We offer a demonstration of cue competition effects in humans under incidental conditions as evidence against the argument that all such effects are based on cognitive inference. The latter supposition would imply that if the necessary information is unavailable to inference then no cue competition should occur. We then discuss the case of unblocking by reinforcer omission, where associative theory predicts an irrational solution to the problem, and consider the phenomenon of the Perruchet effect, in which conscious expectancy and conditioned response dissociate. Further discussion makes use of evidence that people will sometimes provide one solution to a problem when it is presented to them in summary form, and another when they are presented in rapid succession with trial-by trial information. We also demonstrate that people trained on a discrimination may show a peak shift (predicted by associative theory), but given the time and opportunity to detect the relationships between S+ and S-, show rule-based behaviour instead. Finally, we conclude by presenting evidence that research on individual differences suggests that variation in intelligence and explicit problem solving ability are quite unrelated to variation in implicit (associative) learning, and briefly consider the computational implications of our argument, by asking how both associative and propositional processes can be accommodated within a single framework for cognition.

A stimulus-location effect in contingency-governed, but not rule-based, discrimination learning.

“…Nonetheless, rule detection might have been facilitated by instrumental conditioning processes, especially in Condition Replicas. The differences between the two species support the dual-process theory of two separate learning systems (McLaren, Forrest, McLaren et al, 2014;Smith, Boomer, Zakrzewski et al, 2014): pigeons learned via an associative system by detecting contingencies between the presentation of a visual stimulus in a certain location on screen and the occurrence of food reinforcement; humans' behaviour was primarily governed by the deduction and application of task rules (though humans might be able to employ contingencybased learning methods under certain circumstances). Main experiment: Individual human error rates depending on the condition that the participants completed (Replicas or White Keys) and the discrimination task that had to be learned (orientation discrimination or spatial-frequency discrimination).…”

Automaticity and cognitive control in the learned predictiveness effect.

“…The general ideas put forward in this opening paragraph are pervasive in Psychology, taken as fundamental aspects of broad areas of research such as attention (Beck & Kastner, 2009;Desimone & Duncan, 1995;Pashler, 1998), learning McLaren et al, 2014), decision making (Evans, 2008), and perception (Carrasco, 2011). This thesis deals with the interaction between two potent influences on one's behaviour, namely learning and attention.…”

“…Thus, the rate of learning provides an index of attentional change, whereby cue competition is biased in favour of cues with high associability. While many such models remain largely agnostic as to the mechanisms by which associations and therefore associability emerge, it has been widely argued that the formation of associations is automatic (McLaren et al, 1994;McLaren et al, 2014). Thus, to the extent that one assumes that the mechanisms of association are automatic, in that they operate independently of the intentions of the participant, the implicit assumption here is that attentional change is relatively automatic.…”

“…According to this view (e.g., Mackintosh, 2003;McLaren, Green, & Mackintosh, 1994;McLaren et al, 2014;Evans, 2008), learning is the result of both automatic associative processes and higher-order propositional processes operating together, though sometimes in competition (see Le Pelley, Oakshott, & McLaren, 2005). While this position is amenable to characterising the effect of various sources of attentional bias alongside the emergence of learning, much work has traditionally focused on interpretations relating to associability.…”

“…These scores were averaged for predictive and non-predictive cues, as well as novel cues ad and w -z. These are shown separately for Group Standard and Group Novel in Several theorists claim that multiple sources of evidence are integrated in order to judge the causal relationship between co-occurring events, (e.g., McLaren, Green, & Mackintosh, 1994;McLaren et al, 2014). This suggests that there are several possibilities regarding the content of information that is transferred between the two stages of learning in a learned predictiveness procedure.…”