The warning stimulus as retrieval cue: The role of associative memory in temporal preparation

Abstract

In a warned reaction time task, the warning stimulus (S1) initiates a process of temporal preparation, which promotes a speeded response to the impending target stimulus (S2). According to the multiple trace theory of temporal preparation (MTP), participants learn the timing of S2 by storing a memory trace on each trial, which contains a temporal profile of the events on that trial. On each new trial, S1 serves as a retrieval cue that implicitly and associatively activates memory traces created on earlier trials, which jointly drive temporal preparation for S2. The idea that S1 assumes this role as a retrieval cue was tested across eight experiments, in which two different S1s were associated with two different distributions of S1-S2 intervals: one with predominantly short and one with predominantly long intervals. Experiments differed regarding the S1 features that made up a pair, ranging from highly distinct (e.g., tone and flash) to more similar (e.g., red and green flash) and verbal (i.e., “short” vs “long”). Exclusively for pairs of highly distinct S1s, the results showed that the S1 cue modified temporal preparation, even in participants who showed no awareness of the contingency. This cueing effect persisted in a subsequent transfer phase, in which the contingency between S1 and the timing of S2 was broken – a fact participants were informed of in advance. Together, these findings support the role of S1 as an implicit retrieval cue, consistent with MTP.

Introduction

In experiments on human information processing, a trial commonly starts with a neutral warning stimulus followed by a target stimulus to which the participant is instructed to respond. Many researchers adopt this schema for pragmatic reasons. The warning stimulus serves as a time marker that enhances the participant’s temporal preparation for the impending target, which in turn reduces noise in the behavioral data. When asked about the underlying mechanism of temporal preparation, most researchers would probably frame their answer in terms of maximizing alertness for processing the upcoming target or mobilizing resources for action at the expected moment. While such broad notions are not necessarily wrong, they are severely underspecified and fail to provide insight into the underlying processing dynamics.

In the present study, we provide evidence for a role of the warning stimulus as a retrieval cue – a view that is both more detailed and more compelling than prevailing notions. Our view is couched in multiple trace conceptions of long-term memory, which assume that experiences are stored as separate episodic memory traces (e.g., Hintzman, 1986, Logan, 1988, Logan, 1990, Nosofsky and Palmeri, 1997, Nosofsky and Palmeri, 2015). A recent elaboration of this idea is that the warning stimulus retrieves memory traces associated with it, which in turn jointly prepare the organism for the impending target stimulus. Below, we first provide the details of this multiple trace theory of temporal preparation (MTP; Los et al., 2014, Los et al., 2017) and show how it accounts for classic findings in human temporal preparation. Then we return to the idea that the warning stimulus acts as a retrieval cue, which leads up to the empirical contribution of this article.

Temporal preparation has been extensively studied in the variable-foreperiod paradigm. In this paradigm, the researcher varies, within a block of trials, the duration of the foreperiod between a warning stimulus (S1) and a target stimulus (S2), and measures the participant’s response time (RT) with respect to S2. The classical finding is that, as the foreperiod increases, mean RT decreases toward an asymptote (e.g., Niemi and Näätänen, 1981, Woodrow, 1914), indicating a gradual growth of temporal preparation toward a maximum. Furthermore, whereas the RT – foreperiod function is relatively impervious to several experimental manipulations, such as the task set with respect to S2 (e.g., simple versus choice; Bertelson and Boons, 1960, Frith and Done, 1986, Steinborn and Langner, 2012) or the modality and intensity of S1 or S2 (e.g., Grabenhorst et al., 2019, Los and Van der Burg, 2013), it is strongly modified by the distribution of foreperiods. When the distribution of foreperiods is varied in different blocks of trials from negatively skewed (a preponderance of long foreperiods) via uniform to positively skewed (a preponderance of short foreperiods), the RT – foreperiod function becomes progressively less steep, while maintaining a stable asymptote (e.g., Baumeister and Joubert, 1969, Cravo et al., 2017). A benchmark condition is provided by the exponential (“nonageing”) distribution, in which the frequency of consecutive foreperiods decreases according to a fixed rate (e.g., 8:4:2:1), and the RT – foreperiod function has been shown to be approximately flat (e.g., Los et al., 2017, Näätänen, 1970, Näätänen, 1971, Trillenberg et al., 2000).

To account for these and related findings, Los, Kruijne, and Meeter (2014) proposed MTP, which makes three main assumptions. The first assumption concerns within-trial processing dynamics. It holds that the detection of S1 prompts a preactivation of task relevant effectors, which is counteracted throughout the foreperiod by a process of continuous inhibition. Inhibition is lifted when S2 is presented, allowing activation to drive response execution (e.g., Los, 1996, Näätänen, 1971, Narayanan et al., 2006). A large and diverse body of evidence supports this point of view (e.g., Dankner et al., 2017, Los, 2013, Narayanan and Laubach, 2006, Olmos-Solis et al., 2017, Pavlov, 1927, Prut and Fetz, 1999, Toda et al., 2017). For instance, when transcranial magnetic stimulation is applied to human motor cortex, the motor evoked potential measured at the corresponding effector has been shown to be smaller during the foreperiod than at baseline, prior to S1 onset (Davranche et al., 2007, Duque and Ivry, 2009, Hasbroucq et al., 1999). Since this reduced activation has been found for all potential effectors in a choice reaction task, it has been argued to reflect a general mechanism of impulse control that prevents premature response (Davranche et al., 2007, Duque and Ivry, 2009, Jahfari et al., 2010, Prut and Fetz, 1999; for alternative interpretations, see Duque et al., 2017, Hasbroucq et al., 1999).

The second assumption is trace formation. It holds that a unique memory trace is created on each trial, which contains the temporal profile of inhibition (during the foreperiod) and activation (after S2 occurrence) experienced on that trial, along with representations of S1, S2, and the response to S2. As in other multiple trace theories, each new memory trace is added to an accumulating pool of memory traces created on earlier trials (e.g., Hintzman, 1986, Logan, 1988). Furthermore, these traces vary in strength. The strength of each trace is maximal upon its formation and gradually reduces toward an asymptotic value as it grows older (cf. Donkin and Nosofsky, 2012, Howard et al., 2015, Taatgen and Van Rijn, 2011, Wixted, 2004). At this stage of development, we do not commit to any mechanism of strength reduction, be it decay as a function of time (e.g., Baddeley et al., 1975, Barrouillet et al., 2004, Hommel and Frings, 2020) or interference induced by new experiences (e.g., Oberauer et al., 2012, Polyn et al., 2009). Either mechanism would result in the gradual weight reduction we postulate.

The third assumption of MTP is trace expression. It holds that previously formed memory traces jointly determine the state of temporal preparation during the ongoing foreperiod. This process is initiated on each trial by the presentation of S1, which directly and simultaneously retrieves memory traces that contain a corresponding representation of S1 (e.g., Hintzman, 1986, Logan, 1988, Medin and Schaffer, 1978, Ratcliff, 1978). Next, as the foreperiod elapses, each retrieved trace contributes to preparation in accordance with its strength and its momentary value of activation or inhibition. Specifically, at each moment during the foreperiod the state of preparation is determined by the ratio of the weighted activation over inhibition values aggregated across memory traces. Figuratively, trace expression can thus be thought of as a temporal alignment and replaying of previously formed memory traces during the ongoing foreperiod.

Finally, we assume that the state of temporal preparation reached at the moment of S2 presentation determines RT according to an inversely proportional function. This relationship can be appreciated by conceiving the state of temporal preparation as the distance of potential neural excitability relative to a fixed motor-action limit (Näätänen, 1971, Niemi and Näätänen, 1981). However, we refrain from making a principled commitment here because a large body of research has failed to bring clarity on the processing governed by temporal preparation (see Rolke & Ulrich, 2010 for review). Although there seems to be consensus that preparation leaves decision processes unaffected (e.g., Bausenhart et al., 2006, Bertelson and Boons, 1960, Jepma et al., 2009, Los and Schut, 2008), it remains to be resolved whether its main locus of influence concerns perceptual or motor processes (e.g., Correa et al., 2005, Hackley et al., 2007, Mattes and Ulrich, 1997, Müller-Gethmann et al., 2003, Sanders, 1980, Van der Lubbe et al., 2004). In view of this state of affairs, MTP should be evaluated as a learning theory of temporal relationships rather than as a theory of the processes it governs.

Fig. 1 schematically illustrates how MTP explains the classic findings of foreperiod duration, varied at four levels, and foreperiod distribution. In the case of an anti-exponential distribution (Fig. 1A), where the consecutive foreperiods occur with a ratio of 1:2:4:8, temporal preparation is very low just after the presentation of S1 in view of the low ratio of activation over inhibition across memory traces. As time elapses during a long foreperiod, activation gradually takes over, and temporal preparation increases accordingly (Fig. 1C). These dynamics thus give rise to the typically observed steep RT – foreperiod function (Fig. 1E). In the case of an exponential distribution (Fig. 1B), where consecutive foreperiods occur with a ratio of 8:4:2:1, activation starts to dominate inhibition quickly after the presentation of S1. Thus, preparation is already close to ceiling by the time the shortest foreperiod has elapsed and it remains at that level if the foreperiod lengthens (Fig. 1D), yielding the characteristically flat RT – foreperiod function (Fig. 1E).

Before we consider alternative views of temporal preparation, we make a final comment on the activation profile shown in each memory trace of Fig. 1 (in black), which starts prior to the onset of S2 and continues for some time afterwards. This temporal profile can be reproduced by formal models that make use of the property of time cells (e.g., Bakhurin et al., 2017, Eichenbaum, 2014, Mello et al., 2015, Pastalkova et al., 2008, Shankar and Howard, 2012). In these models, each time cell has a unique temporal activation profile, attaining its maximum firing rate at a different moment in time. Thus, when triggered by S1, the population activity of a bank of time cells indicates how much time has elapsed since the presentation of S1. Once S2 is presented, the time cells that are most active at that moment establish a strong association with this event by means of Hebbian learning (e.g., Los et al., 2001, Machado, 1997). Because these time cells exhibit their period of strong activation surrounding the moment of S2 presentation, this results in the corresponding activation profile shown in Fig. 1.

In this section we discuss two alternative views of temporal preparation, one based on the hazard function, the other based on trace conditioning. This exposition allows us to position MTP relative to these alternatives and to show what solutions it offers for otherwise problematic findings.

Explanations based on the hazard function. The hazard function specifies the conditional probability that S2 will be presented at the next possible opportunity during the ongoing foreperiod, given that it has not been presented yet (e.g., Luce, 1986, Nobre et al., 2007, Vangkilde et al., 2012, Vangkilde et al., 2013). It has long been recognized that, for any given foreperiod in any given distribution, mean RT tends to be shorter as hazard is higher, which has led to the wide-spread idea that hazard drives temporal preparation (e.g., Coull, 2009, Cui et al., 2009, Herbst et al., 2018, Janssen and Shadlen, 2005, Niemi and Näätänen, 1981, Trillenberg et al., 2000, Vallesi and Shallice, 2007, Woodrow, 1914). Specifically, the RT – foreperiod functions resulting from different foreperiod distributions, as described in the previous section, are all consistent with the hazard function

It has invariably (though often implicitly) been assumed that the hazard function is derived from the distribution of foreperiods that applies in a block of trials (e.g., Grabenhorst et al., 2019, Janssen and Shadlen, 2005, Trillenberg et al., 2000, Vangkilde et al., 2012). However, this view is problematic because it cannot account for an important class of phenomena based on trial history (Los et al., 2014, Los et al., 2017). Short-term effects of trial history, also called sequential effects, reflect that the RT – foreperiod function is modified by the foreperiods of the immediately preceding trials (Niemi and Näätänen, 1981, Woodrow, 1914; see Los, 2010, for review). Specifically, RT on any trial n has been shown to be relatively long when the foreperiod on that trial is shorter than the foreperiod on trial n – 1 but not when it is equally long or longer (e.g., Capizzi et al., 2015, Drazin, 1961, Langner et al., 2018, Los and Heslenfeld, 2005, Zahn et al., 1963). Earlier trials, at least up to trial n – 2, modify the effect of foreperiod on trial n in a similar way, albeit with much reduced effect size (e.g., Los et al., 2001, Steinborn and Langner, 2012).

Long-term effects of trial history have been demonstrated more recently. Los et al. (2017) presented two groups of participants with either the exponential or the anti-exponential distribution of foreperiods (cf. Fig. 1B and 1A, respectively) in an acquisition phase, and observed the typically flat and steep RT – foreperiod function, respectively. Both groups then received, after explicit instruction, the uniform distribution in a transfer phase of more than 200 trials. The data revealed a flatter RT – foreperiod function throughout the entire transfer phase for the group that had received the exponential distribution during acquisition than for the group that had received the anti-exponential distribution. Extending this finding, Mattiesing, Kruijne, Meeter, and Los (2017) showed that this transfer effect even held when the acquisition and test phase were separated by a full week, attesting to its long-term nature.

Temporal preparation thus comprises a family of phenomena, ranging from short-term (intertrial sequential effects) via medium-term (the effect of foreperiod distribution) to long-term (the transfer effect of foreperiod distribution). In its standard conceptualization, the hazard function is fully determined by the current distribution of foreperiods, so it should drive preparation identically on each trial. Whereas this version suffices to account for medium-term effects, it lacks flexibility to account for the short-term and long-term effects of the family. To accommodate short-term effects, it has been proposed that an additional process operates alongside hazard-driven preparation, resulting in a dual-process model (Vallesi and Shallice, 2007, Vallesi et al., 2013). However, the recent disclosure of long-term effects poses a new challenge to conventional hazard-based models on the other end of the time scale.¹

By contrast, MTP naturally accounts for effects of trial history. It accounts for short-term effects by its assumption that memory traces decrease in strength as they grow older (represented by the thickness of the traces in Fig. 1). Thus, when the foreperiod on trial n is shorter than on trial n – 1, S2 is presented under the strong inhibitory expression of the most recently formed memory trace, which delays response. By contrast, when the foreperiod on trial n is equally long or longer than the foreperiod on trial n – 1, S2 is presented when the inhibitory expression of the most recently formed memory trace is over, thereby leaving response unaffected. MTP also accounts for long-term effects because it denies artificial boundaries between experimental blocks or sessions. As a result, in a transfer phase with a uniform distribution of foreperiods, S1 will still prompt the retrieval of memory traces created during the earlier acquisition phase in which another (e.g., exponential or anti-exponential) distribution applied. Preparation will therefore reveal the influence of these older experiences, even if they occurred a week before, consistent with empirical findings (Mattiesing et al., 2017).

Incidentally, this analysis suggests how a hazard-based approach might be reconciled with the complete family of phenomena in temporal preparation. To do so, the hazard function should be derived, not from the current distribution of foreperiods, but from a memorized distribution of foreperiods that is incrementally updated upon every trial, with a stronger weighting of the most recent trials. In effect, such a ‘memory hazard’ model would stretch hazard to the point that it approaches MTP. But even so, an important difference remains: According to MTP a complete preparatory state emerges from the retrieval of memory traces, whereas a memory hazard view would still need to specify how preparation results from a hazard informed expectancy.

Trace conditioning. To account for short-term effects in temporal preparation, Los, 1996, Los and Van den Heuvel, 2001) proposed a trace-conditioning model, which is in many ways akin to MTP. However, instead of assuming that a new memory trace is created on every trial, this model assumes that a single memory trace is continuously updated by inhibition during the foreperiod (“extinction”) and by activation upon the presentation of S2 (“reinforcement”). A formal version of this model (based on Machado, 1997) accurately accounts for the asymmetry of (higher-order) sequential effects (Los, 2013, Los et al., 2001).

The trace-conditioning model was originally intended as a general account of the phenomena of temporal preparation, based on the idea that medium-term effects would result from a propagation of short-term effects. For instance, favorable short-short sequences occur much more frequently under the exponential than under the uniform distribution, which therefore must contribute to the reduced slope of the RT – foreperiod function under the exponential distribution. However, a direct test of this idea revealed that this propagation of short-term effects is far too modest to account for the full-fledged effect of foreperiod distribution (Los & Agter, 2005; see Vallesi and Shallice, 2007, Vallesi et al., 2013 for additional criticisms). By extension, the more recently disclosed long-term effects are obviously way out of reach of the trace-conditioning model.

MTP corrects the myopia of the trace-conditioning model by assuming that all experiences are stored in separate memory traces. While the higher strength of recent memory traces still allows MTP to account for short-term effects, the relative frequency and longevity of all traces allow it to account for medium-term and long-term effects in addition. Thus, the system is flexible in prioritizing recent information relevant to the task at hand, while being stable by taking into account information that proved successful in the more distant past (see Los et al., 2014, for a comprehensive discussion).

An important insight ensuing from MTP is that S1 serves as a retrieval cue that prompts its associated memory traces to drive temporal preparation. This mechanism hinges on the assumption that a representation of S1 is stored in each newly created memory trace, making it addressable when the same S1 is presented on a future trial.

In the present study, we tested this assumption. We used the variable-foreperiod paradigm, with two different S1 cues, denoted as S1_A and S1_E, and two different distributions of foreperiods, the anti-exponential and exponential distribution. During an acquisition phase, S1_A was associated with the anti-exponential distribution (Fig. 1A), while S1_E was associated with the exponential distribution (Fig. 1B). Participants were not informed of this contingency (with the exception of Experiments 4 and 5). All trials started equiprobably with S1_A or S1_E, followed by a foreperiod randomly sampled from the associated distribution. In the subsequent transfer phase, the contingency between the S1 cue and the distribution of foreperiods was broken, and S1_A and S1_E were both associated with the uniform distribution –a fact participants were explicitly informed of.

According to MTP, the presentation of S1_A or S1_E should lead to the selective retrieval of its associated subset of memory traces. That is, S1_A mainly retrieves memory traces representing long-foreperiod trials whereas S1_E mainly retrieves memory traces representing short-foreperiod trials. As a result, the RT – foreperiod functions for S1_A and S1_E should diverge during the acquisition phase, as these S1 cues retrieve more and more memory traces corresponding to their associated distributions. Since this selective retrieval cannot be undone by explicit instruction, the RT – foreperiod functions should only gradually converge again during the transfer phase, as the newly formed memory traces under the uniform distribution come to dominate.

These predictions of MTP obviously rely on the assumption that the total pool of newly created memory traces is separable in two subsets, which are selectively addressable by S1_A and S1_E. If this condition is not met, there will be a strong overlap between the memory representations linked to both S1 cues, and the resulting RT – foreperiod functions for S1_A and S1_E become indistinguishable. This consideration motivated a further goal of the present study: After a successful demonstration in Experiment 1 that the RT – foreperiod function is modified by S1 cue (S1_A or S1_E), we explored the boundaries of this effect in seven follow-up experiments, in which we used different stimulus features to create the S1-pairs.

We finally note that although the verification of the predicted cueing effect would support MTP, it would not necessarily rule out alternative views on temporal preparation. Specifically, the finding of a cueing effect could lead one to propose a mechanism that allows participants to simultaneously learn two hazard functions, contingent on the presented S1. Alternatively, it could be considered a specific case of discrimination learning, which is a standard finding in the conditioning literature (e.g., McLaren and Mackintosh, 2002, Rescorla and Wagner, 1972). However, these considerations should not obscure the fact that MTP is the only account of temporal preparation to date from which the cueing effect derives as a principled prediction. Other views would need to postulate additional mechanisms to accommodate it.