Cognitive modeling ‘versus’ cognitive neuroscience: Competing approaches or complementary levels of explanation?

Author Note: Preparation of this paper was facilitated by a Discovery Grant from the Australian Research Council and an Australian Professorial Fellowship to the first author and by a Centre of Excellence grant from the Australian Research Council to the second author.

There can be no doubt that neuroscientific approaches to explaining the mind and cognition have gained considerable prominence during the last decade or so, to the point where some analysts have detected a declining influence of purely cognitive theorising (Shallice, 2009). Why has the neuroscientific revolution been so popular? What characterises the approach, and how does it differ from purely cognitive means of theorising?

Although the popularity of cognitive neuroscience is indisputable, there are indications that neuroscientific theorising may sometimes be embraced too readily and too uncritically. For example, in a recent study by Weisberg, Keil, Goodstein, Rawson, and Gray (2008), participants were asked to rate how ‘satisfied’ they were with proposed explanations for scientific phenomena. Explanations were designed to be either good—i.e., they actually explained the phenomena—or poor—i.e., they were constructed to be vacuous restatements of the phenomenon rather than true explanations. Weisberg et al. found that people were generally quite sensitive to the quality of explanations, with one exception: When the explanations were augmented by statements about the brain that did not provide any additional substantive information (e.g., ‘Brain scans indicate that [phenomenon X] happens because of the frontal lobe brain circuitry known to be involved in [phenomenon X]’), then bad explanations—i.e., restatements of a phenomenon such as ‘subjects make more mistakes when they have to [X]’—were no longer rejected. That is, the mere mention of the brain improved the perceived quality of a (normatively) bad explanation. This undue ‘allure’ of references to the brain arose with all but the most skilled participants (i.e., including PhD students but not fully trained neuroscientists). Weisberg et al. suggested that ‘. . . something about seeing neuroscience information may encourage people to believe they have received a scientific explanation when they have not’ (p. 470).

We do not suggest that this single study can call into question the value of cognitive neuroscience; nonetheless, the fact that even fairly skilled readers sometimes endorse poor explanations when they are couched in cognitive‐neuroscientific terms raises the intriguing question as to what exactly it is that makes the cognitive‐neuroscientific approach so conceptually alluring. It is impossible to characterise a long‐standing and enormous research endeavour in a few words; nonetheless, several unique attributes of neuroscientific theorising are sufficiently common to warrant exploration.

TAXONOMY OF MIND

First, cognitive neuroscience is often grounded in a ‘taxonomic’ approach to the mind, whereby mental phenomena are ascribed to different cognitive ‘systems’ that are often taken to be anatomically distinct. Those cognitive systems may variously refer to memory stores (Squire, 2009), decision‐making processes (Ferreira, Garcia‐Marques, Sherman, & Sherman, 2006), or modules of reasoning (Sloman, 1996). For example, theorists of categorisation often invoke different cognitive memory systems for ‘rule‐based’ tasks and for ‘information‐integration’ tasks (e.g., Ashby & O'Brien, 2005). The former tasks afford easy verbalisation of rules (e.g., ‘square shapes are in category A, round shapes in category B’), whereas the latter are assumed to defy verbalisation and can only be solved by integrating partial perceptual information across different dimensions (e.g., when Gabor patches differ in frequency and orientation, and both dimensions are relevant for category assignment). Rule‐based tasks are thought to be the domain of an explicit memory system, whereas information‐integration tasks require an implicit or procedural system that is not subject to conscious awareness. This view of categorisation has been extremely successful and has exerted great theoretical appeal. However, the taxonomy is not without its critics: The article by Newell (2012) argues that much of the support for this view has come from work that considers only neuroimaging evidence and that when one also considers the existing behavioural and neuropsychological evidence, it is much less clear that this view has substantial empirical support.

Similarly, the article by Kalish and Dunn (2012) considers another taxonomy; namely, the idea that episodic recognition is based on two qualitatively different forms of evidence, usually described as recollection and familiarity. This is ‘dual‐process theory’; an alternative theory of episodic recognition, ‘single‐process theory’, also exists. These theories can be construed as constitutive theories or as causal theories. Kalish and Dunn argue that when these theories are construed as constitutive theories, cognitive neuroscience has no role to play in the evaluation of the theories. And they also argue that when the theories are instead construed as causal theories, evidence from cognitive neuroscience can be brought to bear on evaluation of the theories only if each theory makes prior commitments as to the neural bases of each of its postulated cognitive components.

Those re‐examinations point to a more general underlying problem with the ‘taxonomic’ approach to cognition; namely, the fact that any unexpected or inconvenient finding can always be accommodated by an increasingly fine‐grained classification of the presumed underlying ‘systems’. For example, in recognition memory, four distinct types of associative novelty (or familiarity) signals have been identified to date (e.g., Düzel, Habib, Guderian, & Heinze, 2004), raising the possibility that these fine distinctions may ultimately turn out to be unwieldy (see Lewandowsky, Ecker, Farrell, & Brown, 2012, for further consideration of this problem).

WEAK SUPPORT FOR CONSTRUCTS

Second, in common with those working in other branches of cognitive science, cognitive neuroscientists often invoke unifying but ‘invisible’ constructs to explain a variety of signature results. For example, there is a near consensus in the cognitive‐neuroscientific literature that memories, once encoded, undergo a process of ‘consolidation’ during which they become more resistant to forgetting (e.g., Wixted, 2004). This consensus is based on strong evidence from a number of arenas, foremost among them sleep research and research involving transcranial magnetic stimulation (e.g., Born, 2010; Marshall, Helgadottir, Mölle, & Born, 2006). Nonetheless, the behavioural identification of consolidation can remain inconclusive because it is, by definition, a compensatory process that stands in opposition to forgetting. In consequence, whatever the observed rate of forgetting in an experiment, it can always be presumed that more forgetting would have occurred without consolidation or that less forgetting would have been observed if consolidation had not been disrupted.

Intriguingly, the notion of consolidation plays little or no role in most contemporary computational models of memory (e.g., Brown & Lewandowsky, 2010), suggesting that consolidation is not a necessary construct to explain the existing large data base on human memory. The article by Lewandowsky et al. (2012) examines this divergence between cognitive‐neuroscientific and cognitive‐computational theorising. They conclude that at least some of the purely behavioural data put forward in support of consolidation are not in fact uniquely supportive of that concept. In other words, although there are numerous results that are compatible with the notion of consolidation if its existence is taken for granted, the construct is not uniquely implicated by the behavioural data that are often interpreted in its support (see also Brown, 2012, for further analysis of this common logical fallacy).

TOWARDS A SYNERGY

Notwithstanding those critical analyses, other articles in this special issue highlight contrasting cases in which a synergy between cognitive‐psychological and cognitive‐neuroscientific approaches appears to be at hand. For example, the article by Brown (2012) provides a particularly nuanced analysis of how functional magnetic resonance imaging data can be used to augment and constrain cognitive theorising. Brown provides a specific example that demonstrated how decision models can be usefully constrained by integrating behavioural and neural data.

Similarly, de Zubicaray (2012) explores a ‘strong inference’ method for testing alternate hypotheses from competing cognitive theories with neuroimaging data. Strong inference relies on observing spatially or chronometrically differentiable activation in response to manipulation of a variable that is thought to affect a specific processing stage or processor. Because this technique is strongly theory‐dependent, de Zubicaray argues that it can more incisively differentiate between competing models than the currently popular large‐scale data‐basing efforts in which functional associations are made without reference to existing modular information processing models (though see Brown, 2012, for a possibly contrary position). de Zubicaray provides compelling examples of the approach using models of spoken word production.

Coltheart (2012) argues cogently for the necessity of having a cognitive level of explanation via demonstrations of the simple point that the literature of experimental cognitive psychology contains many examples of empirical behavioural phenomena for which satisfactory explanations at the cognitive level are available, while no explanations at the neural level are available (because there is no knowledge about how the brain works that is relevant to any of these observed behavioural phenomena).

Finally, Perfors (2012) provides a Bayesian theoretical perspective on hypothesis generation and how the ‘newness’ of a field may determine which hypothesis‐testing approach one pursues. Perfors shows that sparse hypotheses—i.e., those that predict only a few specific outcomes—are best compared by positive tests, that is, by testing the precise prediction of one's favoured hypothesis rather than by seeking its falsification. However, as hypotheses get eliminated, negative tests may achieve more clarification than positive tests. Perfors suggests that the differential theoretical preferences in cognitive neuroscience and cognitive psychology may reflect the differing degrees of maturity of the two fields.

CONCLUSION

The articles in this special issue highlight both divergences and synergies between the different levels of explanation for behavioural phenomena offered by purely cognitive and purely neuroscientific approaches. They also provide critical analysis of some aspect of cognitive neuroscience that, on balance, appear to be taken for granted too readily—in particular, the articles highlight some of the dangers inherent in a taxonomic approach in explaining cognition. Other articles point to the strengths of the cognitive‐neuroscientific approach and to how it can complement purely cognitive theorising. We suggest that the ultimate contribution of this special issue will be to bring the indubitable strengths of cognitive neuroscience into clearer focus by shedding a critical light on those instances in which cognitive‐neuroscientific theorising may have been too enthusiastic or adventurous.

Notes

Author Note: Preparation of this paper was facilitated by a Discovery Grant from the Australian Research Council and an Australian Professorial Fellowship to the first author and by a Centre of Excellence grant from the Australian Research Council to the second author.