Investigating the formation and consolidation of incidentally learned trust.

Sleep and Memory Consolidation

Memory decay is reduced across sleep, suggesting that sleep facilitates some forms of consolidation (i.e., the process by which initially labile memory traces become strong and enduring representations; Rasch & Born, 2013 ). While the majority of studies reporting a memory benefit of sleep have focused on explicitly learned associations ( Cairney, Lindsay, Paller, & Gaskell, 2018 ; Ellenbogen, Hulbert, Stickgold, Dinges, & Thompson-Schill, 2006 ; Gais, Lucas, & Born, 2006 ; Payne et al., 2012 ), other work has demonstrated that implicitly learned information is strengthened across the night ( Durrant, Cairney, & Lewis, 2013 ; Durrant, Taylor, Cairney, & Lewis, 2011 ). Interestingly, sleep has also been shown to support the development of inferential knowledge regarding hierarchical relationships between separate sets of information ( Ellenbogen, Hu, Payne, Titone, & Walker, 2007 ). Yet, whether social inferences pertaining to facial trustworthiness benefit from overnight memory processing is unknown.

The deepest stage of sleep, known as slow-wave sleep (SWS), has been shown to play a particularly important role in consolidating memories formed via both explicit and implicit learning processes ( Durrant et al., 2011 , 2013 ; Gais & Born, 2004 ; Ngo, Martinetz, Born, & Mölle, 2013 ; Ngo et al., 2015 ). According to an influential active systems model ( Born, Rasch, & Gais, 2006 ; Diekelmann & Born, 2010 ; Rasch & Born, 2013 ), memories are reactivated and thereby strengthened in SWS, promoting long-term storage. An experimental technique known as targeted memory reactivation (TMR) has provided compelling evidence for a role of reactivation in overnight consolidation (for reviews, see Cellini & Capuozzo, 2018 ; Schouten, Pereira, Tops, & Louzada, 2017 ). In a typical TMR study, novel information is linked to sounds at encoding; a subset of which are then replayed during SWS in a bid to ‘cue’ the associated memory traces. Memory performance is typically better for cued relative to noncued memories, suggesting that cued representations are selectively reactivated and strengthened during offline periods ( Antony, Gobel, O’Hare, Reber, & Paller, 2012 ; Cousins, El-Deredy, Parkes, Hennies, & Lewis, 2014 ; Oudiette, Antony, Creery, & Paller, 2013 ; Rudoy, Voss, Westerberg, & Paller, 2009 ). Note that TMR effects are typically not observed in wakefulness ( Cairney, Guttesen, El Marj, & Staresina, 2018 ; Rudoy et al., 2009 ; Schönauer, Geisler, & Gais, 2014 ; Schreiner & Rasch, 2015 ), indicating that memory cueing impacts upon mnemonic operations unique to sleep (although see: Oudiette et al., 2013 ; Tambini, Berners-Lee, & Davachi, 2017 ). Recent work has furthermore indicated that TMR can be used to stabilize implicitly learned associations ( Hu et al., 2015 ).

In Experiment 1 of the current study, we employed a gaze-cueing procedure to examine how consolidation intervals of sleep or wakefulness influence the decay of validity-contingent trust learning. Using TMR, we furthermore examined whether incidentally learned trust representations for valid-cueing and invalid-cueing faces could be selectively strengthened during sleep or wake. Finally, to probe the persistence of social inferences, we assessed the retention of trust learning following a 1-week delay. Our hypotheses were as follows: (a) incidentally learned representations of trustworthiness would be better retained across sleep relative to wakefulness, (b) the memory benefits of sleep for trust learning would be amplified for representations that were cued via TMR, and (c) trust learning effects would be preserved 1 week after learning.

Another question that we look to address in this study is whether this trust learning is implicit. We describe this effect as incidental learning, as participants are given explicit instructions to ignore the face and to focus on the target objects. As such, any learning about the identities of the faces is the result of tacit processes, as knowing about the trustworthiness of the faces does not help participants complete the task (categorizing objects). Indeed, the only strategic motivation to learn about the faces in this paradigm would be to inhibit misleading gaze cues from untrustworthy identities, but previous research has shown no evidence of such inhibition ( Bayliss & Tipper, 2006 ; Strachan et al., 2016 ). However, we acknowledge that there is a strong possibility that, while incidental, this learning is not implicit—that is that participants, confronted on every trial with a face that makes valid or invalid gaze cues with 100% reliability, could become aware of the crucial manipulation. As such, in Experiment 2 after gaze cueing is complete, we explain the key manipulation of the study and ask participants to report for each individual face whether it had previously looked toward or away from the target location, to test for explicit awareness of these gaze contingencies.

Stimuli

During gaze cueing participants’ task was to categorize objects that appeared on the screen. Target stimuli for this object categorization task were kitchen and garage object images used in Bayliss and Tipper (2006) . There were 13 unique objects in each category (kitchen/garage), and these appeared in both horizontal orientations (i.e., if the object had a handle it could point to the left or to the right). All stimuli were colored in blue. In total there were 52 individual images used in the experiment. Face stimuli were taken from the Karolinska Directed Emotional Faces (KDEF) stimulus set ( Lundkvist, Flykt, & Öhman, 1998 ) and included 16 images: eight male and eight female. These faces were initially selected by eye from a figure in the online supplementary material of Oosterhof and Todorov (2008) , in which the faces from this set are plotted along six judgment dimensions. The faces used were all taken from the center (1 SD from the intersection of all six dimensions) of this plot, so the faces used in our experiments were, compared with the rest of the KDEF set, as close to neutral trait judgments as possible when posing neutral expressions. However, as previous research has shown that trust learning effects are stronger for smiling than neutral faces ( Bayliss et al., 2009 ; Strachan et al., 2016 ) we used the smiling images for each of the chosen identities.

These faces were split into two sets, which would appear as either 100% valid (always looking toward where the target would appear) or 100% invalid (always looking away) cues in the experiment (counterbalanced across participants). The eyes of each face were manipulated using Adobe Photoshop CS6 to generate faces where the eye gaze was straight ahead, left, or right. Although the original images showed direct gaze, manipulated versions showing direct gaze were generated for the gaze-cueing portion of the experiment so that there would not be a change in sclera texture as a result of the gaze shift. Unaltered images were used for the trustworthiness ratings.

For TMR, each face was also associated with one of two synthetic sounds (A and B, each 1 s in duration) taken from Hu et al. (2015) . The sounds can be downloaded from the following link: https://osf.io/q79gv/ . An equal number of valid and invalid faces were paired with Sound A and Sound B.

The study was run on an Intel Core i5 PC with a 21.5″ monitor. The experiment was presented using E-Prime 2.0 software with a white background throughout and the resolution set to 1,024 × 768 pixels. Participants sat approximately 60 cm from the display, and during trustworthiness ratings the face stimuli measured 469 × 650 pixels, while during gaze-cueing the face stimuli measured 307 × 461 pixels (these were smaller to account for other images on the screen during gaze-cueing).

Design and procedure

The experimental structure is shown in Figure 1 . Experiment 1 consisted of two sessions, which were separated by 1 week. In the first session, participants completed three types of experimental block: face-sound familiarization, trustworthiness ratings, and gaze-cueing (described in detail below). Participants in the sleep condition then took a 90-min nap, while those in the wake condition completed a time-matched filler task. TMR was administered during this 90-min interval (see below). In the second session, participants returned to the lab and completed a set of trustworthiness ratings.

In order to familiarize participants with the faces, which has been shown to lead to more stable trust learning ( Strachan & Tipper, 2017 ), and to ensure that face-sound associations used for TMR were well established, participants completed a familiarization task at the beginning of the experiment.

In this task, two faces were presented on the left and right side of the screen. Each pair of faces consisted of one identity associated with Sound A and one identity associated with Sound B. Participants then heard one of the sounds (A or B) via headphones and were instructed to respond with which of the two faces that sound was associated (the left face or the right face, using the keys Z and M, respectively). The faces would remain until a decision was made, although the sound would only play once. After the participant had made a decision, they were given feedback on whether they were correct or incorrect. As such, while participants started off by guessing, they learned the sound associations over multiple exposures. Participants completed this task as many times as it took for them to go through one full block (each face presented once on the left, once on the right, making 32 trials) without making a single mistake.

Trustworthiness ratings

Participants made trustworthiness ratings of all 16 faces used in the experiment. They saw the nonmanipulated original face images showing direct gaze in a random order and were instructed to rate them on trustworthiness by clicking on a linear scale with the mouse. At the beginning of each trial, a calibration screen appeared with the question “How TRUSTWORTHY do you think this person is?” with the word “START” written vertically beneath it. Participants clicked the word “START” to begin the trial, after which the face would appear for 1,000 ms. Then the face disappeared and was replaced with an uninterrupted horizontal line rating scale with “−” and “+” at the left and right end of the line, respectively. Participants were told to click at the point along the line that they thought corresponded to how trustworthy the person was, with more trustworthy ratings closer to the “+” label and less trustworthy ratings closer to the “−” label. The x location of the final mouse position was coded as a point between −100 and +100, with 0 being the absolute center of the line on the screen. We used an uninterrupted line scale with no marks or numbers to reduce the chance of participants explicitly remembering the rating they had previously given particular faces and trying to be consistent with their earlier choices.

Participants completed four trustworthiness rating blocks, during which each face was presented once in a random order, resulting in 16 trials per rating block. The first block followed the face-sound familiarization trials but preceded the gaze-cueing portion of the experiment ( preexperiment rating or baseline ). The second block came immediately after the gaze-cueing but before the 90-min sleep/wake interval ( preinterval rating ). The third block followed the interval ( postinterval rating ). The final block took place 1 week after the original session ( 1-week rating ).

Gaze-cueing

During the gaze-cueing portion of the experiment, participants were instructed to respond to object images that appeared on the left or right side of the screen after a face had made a gaze shift (i.e., left of right). Participants were explicitly instructed to ignore the face as it was intended to serve as a distractor. They were instead told to focus on deciding whether the object was a kitchen or garage object, using the assigned keys H and the space bar (counterbalanced mapping across participants).

At the beginning of a trial, a fixation cross appeared on the screen for 600 ms. Following fixation, a face appeared on the screen showing direct gaze for 2,500 ms, during which the associated sound would play. Then the face shifted its gaze to either the left or the right. Five-hundred milliseconds after the gaze shift, the object would then appear in the gaze-cued (valid trial) or gaze-uncued (invalid trial) location. The target object remained on screen for 3000ms or until the participant’s kitchen/garage response was logged. The face then shifted back to direct gaze for 1,000 ms, followed by a 1,000-ms feedback screen which showed “XX” in red for trials where an error was committed (i.e., incorrect kitchen/garage decision). Feedback was only provided on error trials in an effort to reduce posterror slowing ( Compton, Heaton, & Ozer, 2017 ). There was a 500-ms blank display interval between trials.

In total there were seven blocks with 32 trials each, with each face appearing twice in each block, once gazing left and once right. As such, each face appeared 14 times in total throughout the experiment, always producing either valid or invalid cues (depending on the identity).

Sleep and wake delays

Participants in the sleep condition were left to nap in a laboratory bedroom for 90 min while their brain activity was monitored with polysomnography (PSG). An Embla N7000 PSG system with RemLogic 3.4 software was used to monitor sleep. After the scalp was cleaned with NuPrep exfoliating agent (Weaver and Company, Aurora, CO, USA), gold plated electrodes were attached using EC2 electrode cream (Grass Technologies, West Warwick, RI, USA). EEG scalp electrodes were attached according to the international 10–20 system at six locations: frontal (F3, F4), central (C3, C4), and occipital (O1, O2), and each was referenced to the contralateral mastoid. Left and right electrooculography electrodes were attached, as were electromyography (EMG) electrodes at the mentalis and submentalis bilaterally, and a ground electrode was attached to the forehead. Each electrode had a connection impedance of <5 kΩ. All online signals were digitally sampled at 200 Hz. Sleep scoring was carried out in accordance with the criteria of the American Academy of Sleep Medicine ( Berry, Brooks, Gamaldo, Harding, & Vaughn, 2012 ).

TMR was initiated when participants were in the N2/SWS transition. The sound (A or B) was played continuously with a randomized interstimulus interval between 3.5 s and 6.5 s. TMR continued for as long as participants were in SWS, but immediately paused if they showed signs of microarousal, awakening or transition into another sleep stage. TMR was restarted if participants reentered SWS.

To habituate participants to auditory stimulation during sleep, and thus reduce the risk of arousals or awakenings during sound replay, low-intensity Brown noise was played into the bedroom for the entirety of the nap phase. The overall sound intensity (TMR cues + background noise) was ∼50 dB.

Participants in the wake group played the online game Bubbleshooter ( http://shooter-bubble.com ) for the first 30 min of the interval. After this, they completed a working memory task. On each trial, a series of random letters were presented, one after another, in-between valid or invalid sentences, with the series length varying from two to seven letters. Participants were required to retain the series of letters in working memory while making judgments on the validity of the sentences that appeared between each letter, and then report the full series of letters at the end of the trial. At the same time as completing this task, the TMR sound was presented as described above. This approach ensured that participants were sufficiently distracted from the sounds ( Cairney, Guttesen et al., 2018 ), and, thus, would be unlikely to actively retrieve the associated faces. After completing the working memory task, participants played Bubbleshooter again for the remaining 30 min of the interval.

Trust ratings

Trust ratings were recorded in four sessions: preexperiment (before gaze cueing), preinterval (before the 90-min sleep/wake interval), postinterval (immediately following interval), and 1 week . in each session, ratings were between −100 and +100 for each identity used in the experiment.

For the first run of analysis we examined whether incidental trust learning effects were replicated. We included session and validity as independent factors in a repeated measures ANOVA. However, because we were principally interested in the change in learning between later sessions, we transformed the data into trust change scores to remove the influence of baseline ratings from this analysis. For each participant, we first generated mean trustworthiness ratings for each session, for each level of face validity. Using the preexperiment mean ratings as baseline, at each subsequent session we calculated the change in trust from baseline for valid and invalid types of face separately. We submitted these change scores to a 2 × 3 (Validity: Valid, Invalid × Session; Preinterval, Postinterval, 1 Week) repeated measures ANOVA to see if there was decay in trust learning over time. For the three sessions’ change scores we ran separate follow-up Bonferroni-corrected t tests to test whether trust learning was significant at each time point.

When analyzing the data in terms of the effect of sleep, we were particularly interested in how sleep (and TMR) affected the magnitude of learned trustworthiness judgments, both between the pre- and postinterval sessions, and across the week-long interval (relative to the postinterval session).

To address this question, we first calculated a trust learning index at each time point (session) by subtracting invalid change scores from valid changes scores (as such, a greater value indicated a more extreme distinction in trust scores based on the gaze validity of the faces). Next, we calculated a trust change index for the postinterval session ( t = 3) and the 1-week session ( t = 4) using Formula 1 separately for TMR-cued and noncued faces (TMR-on and TMR-off) in the sleep and wake groups. These indices were then compared in separate (one per session of interest) 2 × 2 mixed ANOVAs with group (sleep/wake) as a between-subjects factor and TMR (on/off) as a within-subjects factor. ( V t − I t ) − ( V t − 1 − I t − 1 ) = trust change index

Formula 1

Formula to calculate a trust change index for the third and fourth trust rating sessions. V and I indicate baseline-adjusted trustworthiness ratings to valid and invalid faces, respectively. t refers to the session of interest ( t = 3: postinterval session; t = 4: 1-week session).

All trust ratings were again analyzed using the ez package in R. Wherever Mauchly’s assumption of sphericity was violated we used a Greenhouse Geisser correction of the degrees of freedom.

Null results that were considered theoretically important were followed up with Bayesian analyses to evaluate the evidence in support of the null. This Bayesian analysis involved calculating a Bayes factor indicating the probability of the observed data under the null hypothesis (H0) relative to the alternative hypothesis (H1; BF 01 ). This allows for a clearer interpretation of the data. For example, BF 01 = 3 indicates that the data are three times more likely under H0 than H1. Bayesian information criterion probabilities (pBIC) were also calculated as these provide a graded level of evidence regarding which model (H1 or H0) is more strongly supported, given the data. These values are calculated using an approach that applies simple transformations to the sum of squares of the frequentist ANOVA, as outlined in Masson (2011) .