In this target article, Mitchell et al. argue for a propositional account of human learning, rather than a dual-process model that allows for propositional and associative (what they call the “link model”) processes to operate concurrently. The issue at hand, then, is whether we need to postulate associative processes in addition to propositional ones; the converse argument, whether we need to postulate propositional processes in addition to associative processes, can be left for another time. But let me be quite clear: I am of the view that we need to appeal to both if we are to understand learning in humans.

The approach taken in this commentary is to point out differences in learning and performance under conditions that should favour either propositional or associative learning. Mitchell et al. consider a number of these cases, but perhaps do not do them justice. I take as my first example their review of the Le Pelley et al. (Reference Le Pelley, Oakeshott and McLaren2005a) demonstration of unblocking in humans. In these experiments, it was demonstrated that a design such as A−>O ₁ followed by O ₂|AB−> O₁ followed by O ₃ revealed that learning to B was greater for outcome 1 (O ₁) than in a conventional blocking design where the second phase had the compound followed by O ₁ then O ₂. This finding was predicted on the basis of Mackintosh's (1975) associative theory of learning, which has received experimental support in animals other than human. To dismiss it by saying that it is possible that in a complex design the human participants had forgotten earlier trials and knew something had changed but were not sure whether it was O ₁ or O ₂, ignores this background. As an explanation of the phenomenon, it is terrifically weak. We are expected to allow that propositional learning and an automatic memory (that is definitely not associative?) are both imperfect, and so people make mistakes, which just happen to be the ones that associative theories predict.

This does sound rather implausible, and it is, even though the authors reassure us that it can be tested. Their proposal is to make the outcomes more distinctive, thus reducing any confusion between them, and so the effect (unblocking) should go away. In fact, if the outcomes were made that distinct from one another, the same associative theory that predicted the original result would now predict that the effect would go away as well, as the alpha change that leads to unblocking is to some extent reinforcer-specific in this model. This result has also been found in humans in another experiment by Le Pelley and colleagues (Le Pelley et al. Reference Le Pelley, Oakeshott, Wills and McLaren2005b), in which changing outcomes from those that are generally “nice” to those that are generally “nasty” (and vice versa) prevented alpha effects that were generated by manipulating the predictiveness of certain cues during training. So we are left with a “test” of their account that fails to distinguish between it and the very associative theory that motivated the experiment in the first place. Not much of a test!

Mitchell et al. also fail to take into consideration a number of other studies that demonstrate a different pattern of results when learning is dominated by either rule-based (hence propositional) or associative processes. People show a peak shift, like pigeons, when they are tested on a dimension after relatively little experience with it, and when they are unable to verbalise any rule that captures the discrimination (Jones & McLaren Reference Jones, McLaren, Hahn and Stoness1999; and see Livesey & McLaren, forthcoming). This pattern of responding changes (to a monotonic function across the dimension) after extensive experience with the stimuli and when people can verbalise the correct rule. In the spirit of the target article, I would expect the response to be that this does not demonstrate associative learning, but instead, incorrect rule induction or imperfect application of a rule in some way. If this characterisation of Mitchell et al.'s position is right, then it is impossible to defend against. There will always, with sufficient ingenuity, be some incorrect or imperfect rule that can be appealed to that fits the behaviour. But it simply ignores the fact that these studies were based on predictions made before the fact by a dual-process model, not after the data had been collected. Any sufficiently complete learning system can explain any pattern of results once the pattern is known; the trick is to predict them in advance. The “imperfect rule” approach will never do that.

My final example, which reinforces the point just made, concerns work on sequence learning done by myself and Rainer Spiegel (Spiegel & McLaren Reference Spiegel and McLaren2006). In this series of experiments, we show that the predictions made by the Simple Recurrent Network (SRN), even though at times quite counter-intuitive at first sight, are nevertheless borne out by the experimental results obtained with humans. For example, training on sequences of the form ABC … CBA and ABBC … CBBA where C … C can be one, three, or five C terms, leads to the ability to respond faster and more accurately to the term after the first B following the C terms, as it is predicted by the rule “The number of Bs after the Cs is the same as the number experienced before the Cs.” However, this rule would not predict that when tested after acquisition involving sequences with an odd number of C elements on a sequence, such as ABCCCCBA, the result would be that no learning was displayed, in that responding to the A after the Cs was not facilitated. This was predicted by the SRN, and was the case in our experiments. Many other counter-intuitive effects are reported in Spiegel and McLaren (Reference Spiegel and McLaren2006) that all closely follow the predictions made by this associative model. I am not sure whether Mitchell et al. will opt for the “wrong or imperfect rule” approach here or simply rule these data out of court on the grounds that the SRN is not a standard model of associative learning, but either way the purely propositional account seems implausible to the point of incredulity when confronted by these data.

The target article is an enjoyable attempt to make the case for an exclusively propositional account of human learning, but I fully expect that the learning theories of tomorrow will make use of both rules and associations, and that the attempt to restrict theorising to one or the other will quickly pass.