Stimulus–response bindings in priming

1Our daily lives involve interacting with a large number of stimuli. Many of these stimuli occur not only once, but recur at different timescales. We therefore need to learn rapidly how to process these recurring stimuli, without necessarily intentionally recalling prior experiences with those stimuli. One example of this rapid learning is priming, in which a change in the mental processing of a stimulus is normally measured by an overt behavioral response cued by that stimulus (Box 1). Priming is often interpreted as facilitation of one or more of the computations, or ‘component processes’, that are necessary to generate that response. In a typical laboratory experiment, for example, participants might press one of two buttons depending on whether they judge a visually presented object to be living or nonliving, for which priming would be apparent if their average reaction time (RT) for this judgment is shorter for repeated than for initial presentations of the objects. In this example, the component processes that are facilitated might include perceptual identification of the object depicted (perceptual priming) and/or retrieval of semantic information about that object (conceptual priming).

2However, it has long been suspected that priming can also result from directly associating, or binding, a stimulus and response. If these S–R bindings are retrieved when the stimulus is repeated, the response can be produced without necessarily recapitulating the component processes that were initially used to generate that response (Figure 1A). Evidence for S–R bindings has been found in all major priming paradigms (Box 1 and Figure 1B): repetition priming, negative priming, and masked priming.

3The study of S–R bindings is important for two reasons. First, the presence of S–R bindings potentially confounds interpretation of many priming effects; for example, whether unattended items are truly inhibited or whether semantic representations can really be accessed unconsciously. More importantly, S–R bindings are now recognized to play a vital role in the control of action; a role that goes beyond the mere acceleration of stimulus-driven action. For example, research reviewed below indicates that S–R bindings encode information at multiple levels of abstraction, furnishing considerable flexibility and context sensitivity in their deployment. However, despite the recent interest in S–R bindings, many crucial questions remain unanswered.

4Although the concept of S–R bindings is not new (Box 2), a recent resurgence in interest has been driven by evidence for their greater prevalence and flexibility than previously thought. For example, S–R bindings are far from being simple associations between a specific stimulus and specific response; rather, they appear to be structured bindings involving multiple levels of representation of responses, stimuli, and tasks (Figure 2). Moreover, these bindings do not need to be gradually learned; they can be formed from a single pairing of a stimulus and response.

5Despite evidence for effector-specific response codes, there is general agreement that responses can also be represented by their goal, rather than just specific motor programs. Priming effects that survive a switch in effector between prime and probe are consistent with this claim (e.g., in the case of negative priming).

6However, priming that is invariant to a change in effector is not necessarily evidence for abstract response representations in S–R bindings, because the residual priming could reflect other factors, like facilitation of component processes (or inhibition of a stimulus representation in the case of ). To identify the level of response representation within an S–R binding, one must measure the difference in priming between a congruent condition (where the same response is given each time a stimulus is presented) and an incongruent condition (where the opposite response is given) or an unrelated condition (where a different type of response is given). In the case of repetition priming, for example, Horner and Henson showed a reduction in priming when the response to a repeated stimulus switched from a vocal yes/no to a manual key press (relative to the congruent condition of a manual response to both prime and probe). This suggests that the specific motor action is indeed encoded in the S–R binding. However, priming from a vocal yes/no response to manual response was still greater than when the prime object was simply named. This suggests that S–R bindings additionally encode more abstract response representations, such as a yes/no decision (see also ).

7Further research has suggested that S–R bindings can encode even more abstract response representations, such as the particular classification made (such as living versus nonliving). By changing the reference object during a relative-size judgment task (an experimental manipulation introduced by ), Horner and Henson found greater repetition priming for objects that maintained the same bigger/smaller classification despite this reference change than for objects for which the classification changed. In the latter case, note that the yes/no decision and motor action were unchanged; only the classification label changed. These data suggest that S–R bindings can simultaneously encode at least three levels of response representation: action, decision, and classification (Figure 2). Furthermore, response representations like these have been shown to have independent effects on behavior.

8It has also been suggested that stimuli can be bound to nonspecific ‘stop codes’ that, when retrieved, inhibit responses in any ongoing task. It has even been proposed that stimuli can be bound to attentional filters that have been previously applied to those stimuli. This has been used to explain stimulus-specific congruency-proportion effects in Stroop tasks, where congruency effects are larger for stimuli that have been presented in contexts with a higher proportion of other congruent stimuli.

9Multiple simultaneous levels of response representation potentially allow rapid execution of a specific action, as appropriate, for example, if the context (e.g., task) is unchanged, as well as allowing more flexible response options if the context changes. The downside of this flexibility for priming experiments is that simply changing the effector between prime and probe is not sufficient to ‘control for’ S–R bindings. Likewise, the presence of more specific response codes means that changing tasks between prime and probe is not a sufficient control either, if both tasks require a yes/no decision or the same motor action. To properly investigate priming irrespective of the influence of S–R bindings, all levels of response representation must be controlled simultaneously.

10Similar questions relate to the nature of stimulus representations within S–R bindings. Priming, and its modulation by response congruency, has been shown to decrease with decreasing perceptual similarity between prime and probe, consistent with S–R bindings encoding relatively form-specific representations. However, response congruency effects have also been found despite switching from object pictures to written object names or from object pictures to object sounds. This again suggests that S–R bindings can encode multiple levels of stimulus representation, including at the abstract level of stimulus ‘identity’ (Figure 2).

11Response congruency effects have been shown for semantically related stimuli. This raises the question of whether bindings can be formed between responses and the features that comprise stimuli (see ). In the case of masked priming, for example, such ‘feature–response’ (F–R) bindings may explain priming from stimuli that occur only once in an experiment: so-called ‘novel’ primes. This finding has been assumed to exclude S–R bindings (although see Box 2). If related stimuli have been seen (as probes) and paired with a response, such that features of those stimuli become bound with that response, later repetition of some of those features in a novel (but related) prime stimulus may be sufficient to retrieve the response and hence prime the subsequent probe. This hypothesis is consistent with priming by novel words that comprise parts of words seen previously as probes and with claims that masked priming from novel stimuli occurs only when stimuli come from a small and tightly related stimulus set.

12S–R bindings may also include representations of more than one concurrent stimulus. In negative priming paradigms, for example, there is evidence that the target and distractor stimulus also become bound together, independent of their binding to the response. Such ‘S–S bindings’ seem to be determined by the principles of perceptual grouping. There is also evidence of S–S bindings in associative priming tasks where the response requires comparing two or more concurrent stimuli. Again, these data imply a more complex picture of S–R bindings than is normally conceived, including multiple levels of stimulus as well as response representation, bindings between stimulus features and responses, and bindings between multiple stimuli. This complexity affords yet further flexibility in, for example, allowing learned responses to be triggered not only by the same stimulus, but also by similar stimuli.

13Aspects of the concurrent context might also be bound with the stimulus and response. One example is the task set in which a stimulus is encountered. It has been shown that the typical ‘task-switch cost’, which reflects slower RTs for trials preceded by a different relative to same task, is increased if stimuli are repeated across the tasks. This suggests that the repetition of a stimulus automatically retrieves the previous task set associated with that stimulus, which can interfere with any new task set (also see ). Importantly, Waszak and colleagues argued that S–R bindings are more likely to be retrieved if they were compiled under a task set that remains active during the probe trial (given that a previous task set remains active for a certain time after a task switch: so-called ‘task-set inertia’ ). Task set-dependent retrieval clearly makes adaptive sense, in that one would not want all previous responses that have been associated with a stimulus constantly to compete with current behavioral goals (cf. ‘utilization’ behavior ). Other types of spatial or temporal context (e.g., laboratory setting) may also mediate S–R retrieval. Overlap in this level of context may explain why prior responses can still be cued by a repeated stimulus despite a switch in task, at least when specific response options are shared between the tasks (as in Figure 2).

14More recently, a new line of research has explored how S–R bindings might be formed simply by verbal instruction. For example, Wenke and collaborators presented participants with a set of S–R mappings (e.g., N = left key, K = right key) for one task (Task A). Before attempting Task A, participants performed another task, Task B. Although the instructed mappings for Task A were irrelevant to Task B, they interfered with the performance of Task B when the stimuli in Task B overlapped with those instructed. Instruction-based S–R bindings might allow people quickly to implement any arbitrary S–R mapping and to use this mapping to guide behavior early in the learning of complex skills. However, the nature of instructed S–R mappings is not yet well known. One possibility is that instructions result in covert execution of the instructed mapping, with results similar to the overt application of the mapping.

15Although attention and awareness are intimately related, one can be aware of a stimulus despite it not being the focus of attention, as in negative priming, or one can attend to a specific point in time and space but not be aware of a stimulus presented at that point, as in masked priming. To what extent are attention and awareness important for the encoding and/or retrieval of S–R bindings?

16The negative priming paradigm has shown that attention is not necessary for encoding S–R bindings. For example, Rothermund et al. presented strings of five letters (e.g., BFBFB), in which only the second and fourth were task relevant. They found the standard negative priming effect when the distractor letters (in the other positions) became task relevant (i.e., targets) in a subsequent trial and the correct response was incongruent with that given on the original trial, but positive priming when the response was congruent. Indeed, it may make adaptive sense to bind all stimuli to responses when encoding new experiences, because one does not always know which stimulus will be relevant in the future. Frings and colleagues also found a response congruency effect when distractor stimuli were repeated as distractors, suggesting that attention is not necessary for retrieval of S–R bindings either (see also for evidence from repetition priming). Nonetheless, negative priming experiments using other stimulus configurations or longer lags between repetitions (A. Horner, PhD thesis, University of Cambridge, 2010) suggest that attention may sometimes be necessary. One possibility is that bindings initially occur between all stimuli, attended or unattended, in a short-lived ‘event file’, but only the bindings to attended stimuli last longer.

17Although attention appears to be necessary for masked priming, response congruency effects in masked priming suggest that awareness is not necessary for retrieving S–R bindings. Eckstein and Henson, however, found no evidence of response congruency effects for masked primes that were never seen unmasked, suggesting that awareness is necessary for encoding such bindings. Although other studies have found main effects of priming from primes never seen unmasked, this residual priming could reflect unconscious facilitation of component processes rather than subliminal encoding of S–R (or F–R) bindings. Again, to establish the role of a factor like awareness or attention in the encoding or retrieval of S–R bindings per se, one needs to find an interaction between that factor and response congruency.

18Several questions remain about the nature of S–R bindings and how they interact with other processes to determine behavior. For example, does each pairing of a stimulus and response produce a new S–R binding or progressively strengthen an association between an existing stimulus and response representation? Either possibility can explain why response congruency effects tend to increase with the number of stimulus–response pairings. The finding that the standard deviation as well as the mean of RTs scales with the number of pairings has been used to argue for separate S–R bindings that race independently to produce the response. However, when responses from these two routes are incongruent, extra time seems necessary to resolve this discrepancy, slowing RTs relative to unprimed trials, which suggests that retrieval of S–R bindings interacts with component processes during the final stages of response selection (Box 2).

19The idea that potential responses retrieved from S–R bindings are vetted by a final stage of response selection affords an extra layer of cognitive control that is likely to be important. For example, in situations where strong top-down control is required, it may be possible to bias selection against the responses retrieved from S–R bindings and in favor of responses generated by component processes (for example, when accuracy is emphasized over speed). Thus, although retrieval of S–R bindings may not necessarily require awareness or attention, retrieval is not ‘automatic’, in the sense that it is modulated by contextual factors like task set (reviewed above), Gestalt mechanisms, semantic matching, and feedback. The cognitive control of response selection then provides an extra level of flexibility, which means that, even when retrieved, S–R bindings do not necessarily dominate our behavior in an inflexible manner. However, the details of this response selection stage remain to be established, and would certainly benefit from computational modelling (Box 3) and possibly convergent evidence from neuroscientific data (Figure 3).

20Although we have focused on RTs, incongruent S–R bindings may also lead to increased error rates. In the case of negative priming, multinomial processing models have been used to analyze the probability of erroneous probe responses due to retrieval of the prime response. If a stimulus from the prime episode is repeated in the probe, the probability of responding erroneously with the prime response is significantly increased compared with when no stimulus is repeated. Errors can therefore be understood as failures of component processes to overcome retrieval of S–R bindings in incongruent trials.

21Although we have emphasized the pervasiveness and flexibility of S–R bindings in priming, we should note that there are priming effects that cannot easily be explained by S–R bindings. One example is residual (positive) priming when all obvious levels of response code are reversed between prime and probe, or at least when there is no obvious overlap in response codes between prime and probe. Such cases arise when tasks like naming or perceptual identification are performed, on the prime for example, together with a different (e.g., classification) task performed on the probe (or vice versa). In such cases, each stimulus would be associated with a unique response that is not repeated in the probe task so could not modulate priming. More generally, there is little doubt that prior processing of an intact visual object can modify subsequent perception of a degraded version (e.g., a binarized image, such as the famous Dalmatian dog ), such that the object is clearly seen when primed but not when unprimed, without any overt behavioral response being made. For further arguments about priming effects that are independent of S–R binding, see. Moreover, researchers should be wary of automatically appealing to S–R bindings to explain priming unless there is direct evidence for their existence, such as modulations of the size of the priming effects by response congruency, as described above. Finally, because we have also raised the possibility of F–R and S–S bindings, it may seem that bindings can explain just about any aspect of human behavior (rendering them somewhat vacuous as explanatory concepts). However, we emphasize that S–R bindings are only assumed here to influence behavior in tasks where stimulus-cued responses overlap with previous responses to those stimuli; that is, in situations where there are after-effects of prior experience.

22Although the cognitive revolution dispensed with the behaviorist claim that all behavior can be understood in terms of S–R learning, such associations undoubtedly play a role in many of our behaviors. Importantly, S–R bindings are more than simple associations between a specific percept and motor act; they are complex, structured representations that simultaneously bind multiple levels of stimulus, response, and task representation. Furthermore, S–R bindings can, under certain experimental conditions, be encoded and retrieved in the absence of attention or awareness. This complexity and ubiquity make it difficult to control for S–R bindings when using priming to investigate other theoretical questions. Moreover, S–R bindings are no longer viewed only as a confound; they have become interesting in their own right (Box 4). Indeed, their ability to allow us to interact with our environment rapidly, yet also flexibly, suggests that they constitute a fundamental aspect of human cognition.