There is little doubt that memory, categorization, and reasoning are all central components of human cognition. But traditionally they have been treated as separate components. Each has been studied using different paradigms and very different theoretical models have been developed to explain each activity. This separation extends to the way that students are taught about each topic; memory and reasoning, for example typically appear in different sections of Cognitive Psychology textbooks (Heit and Hayes, 2008).

To some extent this separation is warranted. Understanding the cognitive processes involved in any one of these areas represents a major scientific challenge. Much of our current understanding of human memory, reasoning, and categorization has only been achieved via the development of specialized methods and models in each domain. However, it could be argued that the fractionation of the study of human cognition sometimes obscures deeper commonalities between tasks and processes. Newell (1973) famously pointed out that the proliferation of phenomena studied in experimental psychology and the dichotomous division of psychological processes (e.g., continuous versus discrete representations, analog versus digital processing) can impede theory development and cumulative scientific progress. As a remedy Newell suggested that cognitive scientists should examine how a single processing system may be extended to explain a variety of disparate tasks [see Dale et al. (2009) and Garcia-Marques and Ferreira (2011) for recent, related arguments]. In this spirit therefore, this article and the other papers in this Research Topic examine the possible commonalities (as well as key differences) among these three key areas of cognition.

Of course, ours is not the only attempt to highlight the deeper connections between seemingly disparate cognitive activities. Logan (2002) for example, developed an instance-based theory that explains phenomena in the domains of both attention and memory. Connectionist (e.g., Kinder and Shanks, 2003; Rogers and McClelland, 2004) and Bayesian models of cognition (e.g., Kemp and Tenenbaum, 2009) often emphasize the common processing mechanisms involved in learning and memory or in different types of reasoning. To our knowledge, however, the current work is one of the first attempts to examine the links between the three domains of memory, reasoning, and categorization.

Elsewhere we have discussed in detail the different types of relations that may be found between two of these domains, reasoning, and memory (Heit and Hayes, 2005, 2011; Heit and Rotello, 2005; Heit et al., 2012). One notable relation is that analogous research questions have been posed in each area. For example, in both fields there is a lively and on-going debate concerning the number of core processes that drive people’s judgments. According to two-process accounts of memory (Rotello and Heit, 1999; Yonelinas, 2002; Wixted and Mickes, 2010), recognition judgments are driven by two different psychological mechanisms; a relatively fast and automatic assessment of item familiarity, and a slower, deterministic process based on item recollection. These processes are often invoked to explain functional dissociations between judgments made under instructions to recognize any item that feels familiar (“know” instructions) in contrast to recognizing items for which associated details can be retrieved (“remember” instructions). In contrast “single process” accounts (Wixted and Stretch, 2004; Dougal and Rotello, 2007; Dunn, 2008) suggest that recognition is driven by a single underlying mechanism (assessment of an item’s “memory strength”) and that the distinction between remember and know judgments can be explained by shifts in the criterion for making a response (with a more conservative criterion used for “remember” instructions; Donaldson, 1996). In terms of signal detection theory (SDT), the single-process theory explains the difference between remember and know judgments as a change in response criterion rather than memory sensitivity.

An analogous distinction between dual- and single-process models can be found in research on deductive and inductive reasoning. Descriptively, deduction involves deciding whether an inference necessarily follows from a given set of premises (e.g., if “Birds have property X” then it follows that “Sparrows have property X”). Induction on the other hand, involves assessing the plausibility of an inference given the premises (e.g., if “Sparrows have property X” then it seems likely that other birds share that feature, even though this is not necessarily true). Dual-process accounts (e.g., de Neys, 2006; Evans, 2008; Heit and Rotello, 2010) argue that each type of reasoning draws on qualitatively different cognitive processes. Induction is characterized as fast, intuitive, and heuristic whereas deduction is characterized as slow and making use of deterministic rules similar to those used in symbolic logic. In contrast, single-process models (e.g., Johnson-Laird, 1994; Oaksford and Chater, 2007) assume that both types of reasoning may be the result of a single common process but may differ in their decision criteria (with a more conservative criterion used in deduction; see Rips, 2001).

An important implication of this analogy is that the methods used to compare processing models in one area, such as memory, may be useful in advancing the debate in another area, such as reasoning (cf. Medin et al., 1995, for a related argument regarding parallels between issues in decision making and similarity judgment). In research on memory for example, signal detection methods have been used to evaluate single- and dual-process models of recognition. Several types of SDT patterns are seen as evidence against a single process, such as differences in sensitivity for different types of judgment, differences in the slope of receiver operating characteristic (ROC) curves, and a non-monotonic relationship between the two types of judgments across probe items (Rotello et al., 2004; Wixted and Mickes, 2010).

Heit et al. (2012) have used an analogous signal-detection approach to test single- and dual-process models of reasoning. Adapting the paradigm first used by Rips (2001), participants were asked to evaluate logically valid or invalid arguments under instructions that emphasized either deductive validity or inductive plausibility. A number of factors such as the typicality of the premise item (Heit and Rotello, 2005), the number of premises consistent with the conclusion (Rotello and Heit, 2009) and the similarity of premise and conclusion items (Heit and Rotello, 2010), were found to have a more profound impact on inductive than deductive judgments. Conversely the logical validity of arguments had more of an effect on deduction. Moreover, when people doing deduction were placed under time pressure they showed reduced sensitivity to the logical validity of items and increased sensitivity to premise-conclusion similarity; in other words, under time pressure those doing deduction made decisions that were more consistent with induction (Heit and Rotello, 2010).

Signal detection analyses of these data generally favored a two-process interpretation. Induction and deduction judgments for valid and invalid items were found to differ in sensitivity, not just in response bias or criterion location (Rotello and Heit, 2009; Heit and Rotello, 2010). Confidence ratings for induction and deduction judgments were also used to construct ROC curves. Contrary to single-process predictions, the shapes of these curves also differed for induction and deduction, with significantly more area under the curve found for deduction ROCs, reflecting better discrimination between valid and invalid items. In subsequent modeling Heit and Rotello compared the fit of two signal detection models to the ROC data. The one-dimensional model assumes that the only difference between induction and deduction was in response criterion. The two-dimensional model (see Figure 1) assumes that two orthogonal types of information, roughly “logical correctness” and “consistency with associative knowledge,” contribute to reasoning, with different weighted proportions of each type of information used in deduction and induction. Monte Carlo simulations found that the single-process model failed to capture key trends in the data, such as differences in the sensitivity of induction and deduction judgments and the effects of argument length and item similarity. The data, however, were well explained by the two-dimensional model.

It would be premature to think that these results have settled the debate in favor of two-process models of reasoning (see Lassiter and Goodman, 2012; Stephens and Dunn, 2013 for alternative single-process interpretations of studies comparing induction and deduction). Nevertheless they can be seen as good examples of the empirical innovation and theory development that can flow from drawing analogies between issues and methods in the memory and reasoning literatures.

A further example of a positive yield from the analogy between memory and reasoning comes from applying signal detection methods to the phenomenon of belief bias in deductive reasoning. Belief bias refers to the fact that the ability to discriminate between logically valid and invalid arguments is apparently affected by the consistency of those arguments with background knowledge and beliefs (Klauer et al., 2000). Consider the following two arguments:

These arguments have an identical logical structure. Neither is valid. However, because the conclusion in the first argument is consistent with people’s background beliefs it is frequently endorsed while the second argument is usually rejected (Evans et al., 1983). Likewise people often have more difficulty accepting logically valid arguments with unbelievable than believable conclusions.

The belief bias has often been measured by calculating an “interaction index.” This is the difference between two difference scores: the rate of positive responding to valid unbelievable arguments minus the response rate to invalid unbelievable arguments, and the response rate to valid believable arguments minus the response rate to invalid believable arguments. This index is usually found to be positive, reflecting the more difficult discrimination for arguments with believable conclusions.

For over three decades the interaction index has played a central role in research and theory development in the field of deductive reasoning (Klauer et al., 2000). Recent work (Dube et al., 2010, 2011; Heit and Rotello, 2014) however, suggests that this approach to measuring belief bias is flawed. This method for measuring deductive reasoning closely parallels a common approach to measuring the accuracy of recognition memory; namely the calculation of a “corrected” recognition score where the false alarm rate is subtracted from the hit rate. The validity of this measure in memory research is based on the assumption of a linear relationship between hits and false alarms. However, in memory research this assumption rarely holds (Dube and Rotello, 2012). Likewise, Heit and Rotello’s work reviewed earlier suggests that deductive judgments cannot be accounted for by a linear dimension of argument strength. In response Dube et al. (2010) collected validity judgments and confidence ratings for deductive syllogisms that varied in validity and in the believability of their conclusions. Analysis of ROC curves revealed that the belief basis effect primarily reflects the use of a more liberal response for arguments with believable as compared with unbelievable conclusions. Contrary to the assumptions of many theories of deductive reasoning, there was no evidence that sensitivity in detecting argument validity was greater for unbelievable than believable arguments. With that said, see Trippas et al. (2013) for evidence using this same methodology that there may be some exceptions based on individual differences.

The work reviewed in the previous section suggests that there are strong analogies between the research questions examined in the memory and reasoning literatures. In this section we examine evidence that points to a deeper relationship between these tasks; namely that they share common processing mechanisms.

Much of this work has focused on commonalities between recognition judgments and inductive reasoning judgments (Heit and Hayes, 2011; Hayes et al., 2013; Hayes and Heit, 2013). Descriptively there seem to be good reasons for suspecting that these tasks may share common processes. Each task begins with the encoding of new information about a number of instances or exemplars. In recognition this amounts to memorizing the items presented at study. In induction this involves learning novel properties of study instances. Encoding is followed by a test phase where the goal is either to distinguish studied from new instances (recognition) or to infer which new instances have the same properties as studied items (induction). Each type of judgment is assumed to involve a comparison of the similarity of novel test probes with members of the study set (Heit and Hayes, 2011). Our thesis is that the process of similarity comparison is the same for recognition and induction, but that the tasks differ in how broadly responses are generalized to novel items. The specific goal of recognition decisions (is the probe similar enough to studied exemplars to conclude that it is identical?) leads to a narrower generalization of responding than in induction (is the probe item similar enough to studied exemplars to conclude that it has the same properties?).

Our strategy for testing these ideas was as follows. First we developed a research paradigm where a common study and test set was presented under either recognition or induction instructions, and we examined the empirical relationship between responses to test items in each task. Second, we manipulated a range of independent variables thought to impact the process of similarity comparison (e.g., study exposure time, exposure frequency, inclusion of child participants) and examined their effects on recognition and induction judgments. Third, we developed an exemplar-based model, GENeralization from EXamples (GEN-EX) that assumed a common processing architecture for memory and reasoning, and fitted it to our empirical data. The details of this work have been reported elsewhere (Heit and Hayes, 2011; Hayes et al., 2013; Hayes and Heit, 2013). Here, however, we illustrate this research strategy by summarizing the method and outcomes of a new study examining the relations between recognition and induction.

In this final section we briefly consider evidence which suggests that memory and reasoning may each be linked to a third important cognitive activity, object categorization.

This review has highlighted the commonalities between reasoning and memory, between categorization and reasoning, and between categorization and memory. In terms of reasoning and memory, we have shown that there is a close analogy between some of the key research questions addressed in each area. Hence, methods used to answer these questions in memory research (e.g., analysis of ROC curves) can be usefully applied to address important issues in reasoning.

The links between memory and reasoning go deeper than surface analogy, however. We have shown that performance on inductive reasoning and recognition memory tasks can be explained by a single processing model that assumes each task has a different threshold for generalizing encoded information to novel instances. This conclusion is strikingly similar to the exemplar-based account of the relationship between categorization and recognition (e.g., Nosofsky, 1988; Nosofsky et al., 2012).

We still lack direct evidence of connections among all three tasks; to our knowledge no one has yet collected inductive, recognition, and categorization judgments for the same set of stimuli. Nevertheless, the pattern of empirical links schematically summarized in Figure 3 is highly suggestive. The Figure reinforces the view that there are processing components common to all three tasks. Our review suggests that one such component is the assessment of the total similarity between novel items and experienced items. The outputs of such similarity comparisons guide responding in each task, although the threshold for generalizing to novel items varies parametrically across the three tasks.

One implication of this review is that, as suggested by Newell (1973), we should re-think some of the conventional boundaries within the domain of cognitive science. Dividing cognition into separate domains such as memory, reasoning, categorization, and so on, may have pedagogical value (e.g., when writing textbooks or teaching classes). It may also make the study of performance in each domain more tractable, at least at an early stage of scientific progress. But it is important to keep in mind that the boundaries between cognitive activities are often due to pre-theoretical assumptions and socially constructed conventions (cf., Kuhn, 1996), rather than direct empirical comparisons or attempts to model underlying processes. One of the key aims of this review and the other invited submissions in this Research Topic was to balance the study of fractionated facets of cognition with an acknowledgment that seemingly disparate tasks may often share common component processes (see Hahn and Chater, 1998; Pothos, 2005 for related approaches). Our hope is that this will encourage other researchers to consider and to investigate the relations among different cognitive activities.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.