No evidence that visual impulses enhance the readout of retrieved long-term memory contents from EEG activity

12 The application of multivariate pattern analysis (MVPA) to electroencephalography (EEG) data allows 13 neuroscientists to track neural representations at temporally fine-grained scales. This approach has 14 been leveraged to study the locus and evolution of long -term memory contents in the brain, but a 15 limiting factor is that decoding performance remains low. A key reason for this is that processes like 16 encoding and retrieval are intrinsically dynamic across trials and participants, and this runs in tension 17 with MVPA and other techniques that rely on consistent ly unfolding neural code s to generate 18 predictions about memory contents. The presentation of visually perturbing stimuli may experimentally 19 regularize brain dynamics, making neural codes more stable across measurements to enhance 20 representational readouts. Such enhancements, which have repeatedly been demonstrated in 21 working memory contexts, remain to our knowledge unexplored in long -term memory tasks. In this 22 study, we evaluated whether visual perturbation s—or pings—improve our ability to predict the 23 category of retrieved images from EEG activity during cued recall. Overall, our findings suggest that 24 while pings evoked a prominent neural response , they did not reliably produce improvements in 25 MVPA-based classification across several analyses. We discuss possibilities that could explain these 26 results, including the role of experimental and analy sis parameter choices and mechanistic 27 differences between working and long-term memory. 28 Key words : Long -term memory, MVPA, decoding, EEG, ping, visual impulse , perturbation, brain 29 dynamics 30 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 23, 2024. ; https://doi.org/10.1101/2024.04.19.590215doi: bioRxiv preprint 2

31 A central question in memory research is how the brain retrieve s information stored in long -term 32 memory (LTM) in the service of adaptive behaviour . This research topic has inspired work from a 33 variety of angles , involving different experimental protocols and methods—including neuroimaging 34 modalities. Electroencephalography (EEG) and magnetoencephalography (MEG ) have proven an 35 integral part of this project because they capture brain dynamics on a sub -second resolution. Such 36 granularity is crucial, given that memory retrieval typically unfolds on the order of seconds, with the 37 neural cascades underpinning memory retrieval evolving even faster (Staresina & Wimber, 2019). 38 To study the evolution of retrieved contents in the brain, one widely pursued family of 39 techniques is multivariate pattern analysis (MVPA)—more broadly known as classification or decoding 40 (Haxby et al., 2014; Grootswagers et al., 2017) . These tools extract and upweight signal dimensions 41 that robustly covary with retrieved memory contents, effectively boosting the signal-to-noise ratio of 42 associated neural activity. MVPA has been successfully used to enrich our understanding of memory, 43 including how information is encoded (Fritch et al., 2020; Kragel et al., 2017; Kuhl et al., 2012) , 44 consolidated (Deuker et al., 2013; Maguire, 2014; Schreiner et al., 2021) , and reinstated during 45 memory recall (i.e., pattern completion; Danker & Anderson, 2010; Favila et al., 2020; Rissman & 46 Wagner, 2012; Xue, 2018). 47 Despite such advancements, the decoding of long-term memory contents in electrophysiology 48 data typically remains only slightly above chance, impairing our ability to study the evolution of neural 49 patterns of interest. One reason for this limitation is that memory processes and their associated brain 50 activity are highly dynamic, which results in variable patterns across trials and participants (ter Wal et 51 al., 2021; Madore & Wagner, 2022). Indeed, MVPA and most other EEG-based analyses rely for their 52 robust predictions on the existence of a detectably constant cascade of neural patterns across 53 measurements (van Bree et al., 2022) . This clash between variability in neural processes on the one 54 hand and the constancy assumption of our analyses on the other may cause us to miss 55 representations of interest, or to obtain different results depending on what experimental event we 56 timelock EEG data to (e.g., retrieval cues vs button presses; Linde-Domingo et al., 2019 ). A factor 57 that further hampers our ability to robustly decode representations is that retrieval comes with fainter 58 neural patterns to begin with compared to perception (Favila et al., 2020; Pearson et al., 2015; Favila 59 et al., 2022) . Together, these points invite creative techniques that improve our ability to infer long-60 term memory representations from dynamic brain activity. 61 In this study, we explore a perturbational method that has the potential to mitigate two issues 62 at the same time: low signal fidelity at the level of measurement, and variability in neural processing 63 dynamics. Specifically, in this EEG study we evaluated whether the presentation of a high contrast 64 visual stimulus—henceforth referred to as a “ping” —during LTM retrieval enhances the readout of 65 signatures of retrieved content. In motivating the hypothesis that pings boost the decodability of LTM 66 representations, we buil t directly onto recent successful efforts in the domain of working memory 67 (WM). In that context, pings have been used to enhance the decodability of the orientation (Wolff et 68 al., 2015, 2017, 2020; Ten Oever et al., 2020; Yang et al., 2023) and colour (Kandemir et al., 2023) of 69 objects actively maintained in WM, as well as anticipated target locations (Duncan et al., 2023) . A 70 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 23, 2024. ; https://doi.org/10.1101/2024.04.19.590215doi: bioRxiv preprint 3 preliminary explanation for these findings is that pings induce a robust evoked response that interacts 71 and indeed boosts the footprint of active neural representations, enhancing their SNR (Barbosa et al., 72 2021). Specifically, pings may regularize neural dynamics across trials and participants by producing 73 a phase reset of brain oscillations that coordinate information processing across neuronal 74 populations. In support of this , visual stimuli presented during memory tasks have been shown to 75 reset the phase of low-frequency brain oscillations that are implicated in encoding and retrieval 76 (Rizzuto et al., 2003; Haque et al., 2015; audiovisual stimuli in Cruzat et al., 2021). Thus, by inducing 77 pings at experimentally controlled moments, researchers may gain a level of control over variability in 78 synchronized activity across information-coding neurons, making their dynamics more similar across 79 measurements to improve the predictive power of MVPA. 80 Importantly however, while ping-based methods have been shown to work in WM contexts, to 81 our knowledge it has not been explored whether they generalize to LTM research in which information 82 is retrieved from stored representations. The purpose of this study then, is to systematically explore 83 the possibility that pings can enhance the readout of reactivated long -term memory contents. To this 84 end, we presented participants with pings as memory processes were actively engaged during cued 85 recall, evaluating whether retrieved representations are more robustly discernible after ping onset. On 86 the whole, we find no compelling evidence that pings boost the classification of retrieved image pairs 87 from EEG activity. 88 89

90 Participants 91 We recruited thirty-three volunteers (22 women, Mage = 23.8 years, SDage = 2.6 years, range = 18 to 92 31) with normal or corrected -to-normal vision, and with no history of epileptic attacks o r 93 neuropsychological conditions that could interfere with the examined study effects. The sample size 94 required to derive a reliable effect was estimated based on (Wolff et al., 2017), though our estimation 95 was limited by the fact that all previous work was in a WM context. One participant did not finish the 96 experiment because they were unwell, and following data inspection, two participants were removed 97 because of poor data quality due to a large number of high impedance channels, and one because of 98 stimulus trigger issues. Thus, EEG-based analyses were conducted based on 29 participants. For 99 behavioural analyses, the first four participants were excluded because of missing button press 100 triggers, which, with the further exclusion of the participant who did not complete the experiment, 101 resulted in an analysis of 28 participants (participants with noisy EEG data were included in the 102 behavioural analysis). 103 Participants were informed about the details of the experiment in advance —including its 104 duration, protocol, and methods —but were left naïve with respect to the purpose and hypotheses 105 associated with the presentation of visual pings. Participants provided their written consent, and after 106 the experiment, they were debriefed and given information about the central manipulation and 107 hypothesis upon request, and they were compensated for their time with £9 per volunteered hour. The 108 study was approved by the Ethical committee of the College of Science and Engineering of the 109 University of Glasgow (Application number: 300210113). 110 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 23, 2024. ; https://doi.org/10.1101/2024.04.19.590215doi: bioRxiv preprint 4 111 Stimulus and apparatus 112 The presentation of stimuli was controlled using PsychoPy (version 2021.2.3; Peirce et al., 2019 ) 113 running on Windows 10. Stimuli were presented on a CRT monitor (53.3 cm; 1024 by 768 pixels) 114 operating at a refresh rate of 60 Hz. Participants were seated in a magnetically shielded room in a 115 chinrest 65 cm from the screen , or at an approximately similar distance from the screen outside the 116 chinrest if they experienced discomfort. Throughout the experiment, a fixation cross (with a visual 117 angle of 0.44°) was presented in the centre of a constantly presented grey background (RGB = 128 118 128 128 ; Psycho Py default ). All centrally presented stimuli overrode the fixation dot. The visual 119 impulse (i.e., ping) was a single full-contrast bullseye stimulus presented at the centre of the screen 120 for 200 m illiseconds (ms; w ith a diameter of 13° and 0.31° cycles per degree ). The ping was 121 generated using MATLAB and edited using GIMP (GNU Image Manipulation Program version 122 2.10.32). 123 In the main memory task, participants learned associations between action verbs and images, 124 and were later prompted with the action verb to retrieve the associated image. The action verbs were 125 selected based on usage frequency (largely based on Linde-Domingo et al., 2019 ) and the image 126 stimulus set was a combination of 192 colour images collated across various royalty free databases, 127 including the Bank of Standardized Stimuli (BOSS, Brodeur et al., 2010), and the SUN database (Xiao 128 et al., 2010). The selected 192 images were constructed to follow a nested category structure of three 129 embedded hierarchical levels. At the top level, the set consisted of 96 objects and 96 scenes, which 130 were in turn composed at the middle level of 48 animate and 48 inanimate objects and 48 indoor and 131 48 outdoor scenes. Moving down to the bottom level, each of the middle level categories branched 132 out into 4 categories (e.g., for animate objects: birds, insects, mammals, and marine animals), each of 133 which contained 12 specific instances (e.g., twelve specific birds). We chose this nested hierarchy of 134 stimulus categories because we did not know a priori what dimension of retrieved memories would be 135 effectively decodable, so we included multiple levels of abstraction and chose one level based on pre-136 defined criteria (See Level Selection). The objects were presented on a white square matching in size 137 to scene images ( i.e., the visual degrees of all stimulus categories were 13°). Key presses were 138 registered using a standard QWERTY keyboard. 139 140 Procedure 141 The main experiment consisted of 8 blocks, each with an encoding, distractor, recall, and recognition 142 phase (Fig. 1A). In total, the main experiment lasted between approximately 45 and 6 5 minutes 143 depending on the duration of self-paced breaks and electrode impedance maintenance. Before the 144 main experiment, participants were provided with a practice run that covered each phase using 145 example verbs and images that were not used in the main experiment. A standardized set of verbal 146 instructions were provided to guide participants through the practice run. If the participant reported not 147 understanding the task or if they did not give accurate responses , the practice run and instructions 148 were repeated. Then, the main experiment commenced, throughout which EEG was acquired. At the 149 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 23, 2024. ; https://doi.org/10.1101/2024.04.19.590215doi: bioRxiv preprint 5 start of each experimental phase, a screen was presented with a reminder of the task instructions and 150 required response keys. 151 152 153 Figure 1 . Paradigm and behavioural results . (A) Experimental paradigm. The encoding phase 154 consisted of a word-image pair learning task. This was followed by a distractor task intended to wash 155 out working memory effects. Then, during the critical recall phase, participants were cued with words 156 to retrieve the paired image while visual perturbations (pings) were presented in 75% of trials. In a 157 fourth phase, recognition performance was tested (not displayed). (B) Average performance during 158 the recognition task for trials with and without pings , collapsing across blocks for each participant . 159 Datapoints are individual participants. (C) Average recognition performance per participant (i.e., 160 collapsing blocks). (D) Average recognition performance per block (i.e., collapsing participants) . (E) 161 Average reaction time during encoding for subsequently recognized and forgotten trials, collapsing 162 across blocks. Note: in B, C, and D, the y-axis is truncated due to high recognition performance. 163 164 165 Ping No ping 0.8 0.9 1 recognition [%] 2468 Block number Average per blockAverage per participant Average per condition 51 0 1 52 0 2 5 Participant Encoding EEG Recall Word-image pairs + jump + + jump + Distractor task Reaction time during encoding Recognized Reaction time [ms] Forgotten n=2194 n=46 01 0 0 0 2 0 0 0 3 0 0 0 4 0 0 0 5000 E A DCB .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 23, 2024. ; https://doi.org/10.1101/2024.04.19.590215doi: bioRxiv preprint 6 In the encoding phase, participants learned to build a mental association between action 166 verbs and paired images. First, a verb was presented for 1500 ms (white, OpenSans font). Then, after 167 1000 ms, the associated image was presented until the spacebar was pressed to indicate the 168 association was encoded (with a 6000 ms limit). Then, after a 1000 ms delay, the next verb was 169 presented. During each block’s encoding phase, 10 unique verb -image pairs were learned in one 170 shot. This resulted in 80 encoded pairs across the full experiment , with the images pseudo-randomly 171 selected from the full stimulus set such as to maintain an equal distribution of top -level stimulus 172 categories (40 objects and 40 scenes) and fully random selection over nested middle and bottom 173 levels for each ping and no-ping condition. 174 The distractor phase that followed was included to flush out WM effects. Here, participants 175 performed an odd-even task lasting 20 seconds. A number between 1 and 99 was presented in the 176 centre of the screen (white, OpenSans font), and participants were instructed to press left key for odd 177 numbers, and right key for even numbers. Following a left or right key press, the next number was 178 presented immediately. Participants’ average performance was displayed at the end of the distractor 179 phase, marked as the proportion of correct responses. This data was not further analysed. 180 Next in each block , the recall phase tested our central manipulation of a ping-based visual 181 perturbation. In this phase, participants recalled the learned verb-image associations of the encoding 182 phase. First, one of the ten encoded verbs was presented for 2000 ms , serving as the retrieval cue 183 that prompted recall of the associated image. In 75% of trials, a visual impulse was presented in 184 either of three time bin s: between 500 to 833.33 ms (“early ping”), 833.34 to 1116.67 ms (“middle 185 ping”), or 1116.68 to 1500 ms (“late ping”) after the onset of the retrieval cue , with a uniform 186 distribution of possible ping times within each bin. This window was chosen on the basis that previous 187 research on cued recall paradigms suggests this is the moment of maximum memory reinstatement 188 (Staresina & Wimber, 2019) . In 25% of trials, no visual impulse was presented in order to derive a 189 baseline for statistical hypothesis testing. Participants pressed the left key to indicate that they had 190 forgotten the image associated with the verb cue, or right key to indicate they remembered it. Key 191 presses only resulted in a new trial after 1700 ms following retrieval cue onset (i.e., 200 ms after the 192 latest possible ping). With presses earlier than that , nothing happened . Participants were given a 193 visual indication that key presses were available via disappearance of the retrieval cue (at its offset; 194 2000 ms). During the recall phase, each of the 10 encoded verb -image pairs were tested four times , 195 resulting in 40 recall trials per block, and 320 trials in total , comprising 160 objects and 160 scenes . 196 Within participants, each of the four conditions —early, middle, late, and no ping —were configured to 197 present object and scene images equally often (i.e., the top-level stimulus category), with the nested 198 mid and bottom -level categories randomized. The sequence of presented s timulus level catego ries, 199 pinging conditions, and verb -image pairs was fully randomized within and across blocks to mitigate 200 order effects. For the within block randomization, while the 40 recall trials were fully randomized, we 201 ensured the same pair was never recalled twice in direct succession. 202 Finally, since the cued recall phase only included subjective memory judgments, a recognition 203 phase was included to obtain an objective measure of memory performance for the verb-image pairs. 204 During this two-alternative forced choice task, o ne of the 10 encoded verbs was presented in the 205 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 23, 2024. ; https://doi.org/10.1101/2024.04.19.590215doi: bioRxiv preprint 7 centre of the screen, with two images (visual angle of 7.8 °) presented underneath, one on the left-206 hand, and one on the right-hand side of the screen. Participants chose which of the two images was 207 paired with the central action verb using a left or right key press (with a 5000 ms time limit). The 208 location of the correctly paired image was randomized between the left and right location. The lure 209 image was always another old image from the immediately preceding encoding phase. Each of the 10 210 encoded verb-image pairs was tested once in a random sequence. Note that we designed this study 211 to expend most of the available study time on the recall phase to maximize the statistical power of our 212 main analysis, with the recognition phase serving chiefly as a basic check to ensure participants were 213 not skipping through the experiment without memorizing verb-image pairs. 214 215 EEG acquisition and preprocessing 216 The data was recorded using a 64-channel passive EEG BrainVision system ( BrainAmp MR; Brain 217 Products) with a sampling rate of 1000 Hz. For our recording software we used BrainVision Recorder 218 (Brain Products) . The 64 Ag/AgCl electrodes were positioned in accordance with the extended 219 international 10-20 system. Due to a necessary change in the recording system, a different EEG cap 220 type (EasyCap) was used for participants 1 to 14 (subset 1) and 15 to 33 (subset 2) . In the first 221 subset, the ground electrode was located on the back of the head, below occipital electrode Oz, and 222 two EOG channels were used to monitor eye movements (placed below and next to the eye ; VEOG 223 and HEOG). In the second subset, the ground electrode was on the midline frontal location AFz, and 224 one EOG channel was used to measure eye movements (placed below the eye ; VEOG ). 225 Furthermore, the cap used in the second subset included channels FT9 and FT10. For event related 226 potential analyses, we included only electrodes common to both caps to enable a universal 227 visualization of brain activity . Most electrode impedances were kept below 25 kiloΩ, and electrodes 228 with outlier impedances were removed during preprocessing, with their associated data interpolated 229 (see below). 230 Preprocessing was performed using FieldTrip (Oostenveld et al., 2011) in MATLAB (the 231 MathWorks). First, the continuous EEG data was split up into two datasets: one with all trials epoched 232 relative to retrieval cues, and one with trials epoched relative to pings and no -ping (defined by 233 randomly sampling ping times of the pinged trials, yielding so -called “pseudo-pings”). Put differently, 234 the data was locked once to 𝑡 = 0 defined as the retrieval cue, and once to 𝑡 = 0 defined as the 235 manipulation of interest or a baseline alternative. In both cases, the epoched trials were 4 seconds in 236 duration (-1 to 3 seconds relative to the event of interest). 237 Each dataset was filtered between 0.05 and 80 Hz and downsampled to 250 Hz . Next, bad 238 trials and channels with outlier impedance levels were manually removed via visual inspection . 239 Subsequently, eye movement and muscle artefacts were identified and removed using ICA 240 decomposition, and removed channels were interpolated using spline interpolation (with the FieldTrip 241 function ft_scalpcurrentdensity). Finally, the data was re -referenced using a common average and a 242 Laplacian method (current source density), deriving separate data structures for cue-locked and ping-243 locked analyses. 244 245 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 23, 2024. ; https://doi.org/10.1101/2024.04.19.590215doi: bioRxiv preprint 8 Behavioural analysis 246 The experiment was designed to result in high or even ceiling memory performance in order to obtain 247 a maximal number of successfully remembered trials, and to optimally evaluate the central hypothesis 248 of a ping -induced decodability enhancement . We report objective performance for the memory test 249 conducted in the recognition phase, both across pinging conditions (Fig. 1B), participants (Fig. 1C), 250 and across blocks (Fig. 1D). We also report subjective judgments during the recall phase, quantifying 251 how often participants report remembering versus forgetting the word -image pair. Reaction time (RT) 252 during the recall phase is uninformative, because as described in the Procedure section, the response 253 key was locked until 1700 ms after cue onset, at which point participants likely had already retrieved 254 the associated image (Staresina & Wimber, 2019) . Indeed, participants reported actively waiting for 255 response buttons to become available. Thus, we instead analysed RT during the encoding phase as a 256 function of whether the word -image pair was subsequently recognized or not. These RT data were 257 collapsed across participants and blocks (Fig. 1E). For the proceeding analyses, both subsequently 258 recognized and forgotten trials were included. 259 260 ERP Analysis 261 For the ERP analyses, only channels common to both electrode cap subsets were used . We applied 262 two types of ERP analyses, one locked to (pseudo -)pings and one to retrieval cues. FieldTrip was 263 used to downsample the data to 250 Hz and a band-pass filter between 0.2 and 40 Hz was used. The 264 data was baseline -corrected from -200 ms to 0 ms from events of interest. For ERP traces, we 265 calculated the average activity across posterior channels (C3, C4, P3, P4, O1, O2, Cz, Pz, Oz, CP1, 266 CP2, C1, C2, P1, P2, CP3, CP4, PO3, PO4, PO7, PO8, CPz, POz ). For ERP topographies, we used 267 the 61 channels common to both ERP cap types. We statistically evaluated whether pings resulted in 268 higher amplitude ERPs compared to no -ping trials using non-parametric Monte Carlo permutation 269 tests applied to each channel, correcting for multiple comparisons using Bonferroni correction as 270 implemented in FieldTrip, averaging activity from 200 to 400 ms after pseudo-pings (alpha = 0.05; 105 271 randomizations). 272 273 MVPA analysis 274 For MVPA, all EEG channels available per electrode cap type were used except EOG channels. 275 Depending on the analysis, w e trained and tested either a multi -class LDA using FieldTrip 276 (ft_timelockstatistics), or a binary -class LDA using the MVPA Light toolbox (Treder, 2020) . We 277 classified EEG data re-referenced using a Laplacian transform on the basis that it accentuates local 278 patterns (Kayser & Tenke, 2015) . All classifier analyses were performed on the recall phase, where 279 our main hypothesis could be evaluated. Unless specified otherwise, analyses were carried out on the 280 retrieval cue-locked dataset. We downsampled the data from 250 Hz to 50 Hz by applying a moving 281 average with a window length of 140 ms, moving in steps of 20 ms . During each step, a Gaussian-282 weighted mean was applied in which the centre data sample of the window was multiplied by 1, and 283 the tail samples by 0.15 (FWHM = ~81 ms). In a subsequent step, sample by sample, the data was z-284 scored across channels (i.e., setting every channel to mean = 0 and standard deviation = 1), followed 285 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 23, 2024. ; https://doi.org/10.1101/2024.04.19.590215doi: bioRxiv preprint 9 by training and testing using LDA . To evaluate decoder performance, w e applied k -fold cross 286 validation (5 folds, with 25 repetitions). For binary class decoding, w e used area under the receiver 287 operating characteristics curve (AUC) as a performance metric because it adjusts for class 288 imbalances (Grootswagers et al., 2017; Xie & Qiu, 2007) . For multi-class decoding, where standard 289 AUC is unavailable, we used accuracy and factored in level-specific differences in chance levels. To 290 infer decoding performance values under the null hypothesis , depending on the analysis, we either 291 used no-ping trials or ping trials with shuffled class labels (100 1st-level permutations, each with 3 292 repetitions). All analyses were restricted to the period before button presses were made (i.e., < 2000 293 ms). 294 295 Level selection 296 We used a multi -class LDA on no-ping trials to determine which retrieved stimulus category (top, 297 middle, or bottom level) is most robustly detectable in the data when our main experimental 298 manipulation was not applied. This level was then locked in for subsequent analyses that relate to our 299 key hypothesis of ping -induced decoder enhancement. We selected the level with a high baseline 300 performance to offer a conservative starting point from which we could establish whether pings are a 301 powerful tool to further enhance decodability. However, as we will see in the results, stimulus 302 selection rationales matter minimally because we found no reliable level differences in the no -ping 303 decoder across levels to begin with. For statistics, we performed a Wilcoxon rank sum test comparing 304 the empirical and shuffled decoding performance for each level, in the way described in the next 305 section. 306 307 Main analysis 308 For the statistical analysis of the main hypothesis, we used two-level permutation testing for the ping 309 versus shuffle decodability comparison, and a Wilcoxon ranked sum test for the ping versus no -ping 310 comparison. The former approach, which is based on van Bree et al., 2022 , implemented the 311 following algorithm in pseudo-code—applied window-by-window: 312 1) For each 2 nd-level permutation (105 times): Grab one random window -specific decodability 313 value from the 1st-level distribution of the 25 permutations of each participant and average the 314 result. This yields 105 permuted averages. 315 2) Generate one empirical p-value by calculating the percentile of the average empirical 316 decoding value within the distribution of permuted averages. 317 The latter approach involved taking the Wilcoxon signed-rank test between the distribution of 318 empirical decoder results and 1st-level permutation results across participants . We opted for a 319 Wilcoxon test over cluster -based methods because it makes minimal assumptions about the 320 distribution of decoding results (Wilcoxon, 1945; Grootswagers et al., 2017). For both approaches, we 321 adjusted the resulting p-values across windows for their false discover y rate (FDR). Since the p -322 values are not independent across time, we applied the approach by Benjamini & Yekutieli (2001). 323 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 23, 2024. ; https://doi.org/10.1101/2024.04.19.590215doi: bioRxiv preprint 10 Finally, for ping-locked analyses we restricted statistical analyses between 0 and 500 ms from 324 ping onset. For analyses locked to retrieval cue, we analysed 500 to 2000 ms from cue, which is the 325 approximate range where memory reactivation is maximal (Staresina & Wimber, 2019). 326 327 Condition-relative decoding peaks 328 In addition to our main analysis, we carried out a presumably more sensitive analysis to evaluate the 329 possibility of ping-induced decoding enhancements. We reasoned that even if visual pings do not 330 offer an enhancement of LTM decoding performance that is strong enough to emerge in a direct ping-331 to-no ping or ping -to-shuffle comparison, there could still be a weaker effect that is detectable by 332 factoring in the relative order of decoding peaks across pinging conditions. Specifically, we tested 333 whether trials with an early, middle, and late ping tended to have, respectively, earlier, later, and even 334 later decoding performance peaks. In other words, we tested to what extent decoding peaks captured 335 ping presentation order s (see Linde-Domingo et al., 2019; Mirjalili et al., 2021 for similar peak 336 selection approaches). 337 First, we took every participant’s SOA -specific decod ing time series —early, middle, and 338 late—and extracted one peak (specified below). Then, we calculated a peak order distance (POD) per 339 participant, defined as the absolute serial distance between the order of extracted peaks and true ping 340 presentation order, given by the formula: 341 342 𝑃𝑂𝐷!"!#!"$%&'()*+ = ' 𝑎𝑏𝑠(𝑝𝑒𝑎𝑘 − 𝑡𝑟𝑢𝑒) 343 344 For example, if the decoder peak came first for early ping trials (1 − 1), third for middle pings trials 345 (3 − 2), and second for late ping trials (2 − 3), this would amount to a POD of two. We divided PODs 346 by the maximum distance (4), normalizing the score between zero and one: 347 348 𝑃𝑂𝐷 = ∑ 𝑎𝑏𝑠(𝑝𝑒𝑎𝑘 − 𝑡𝑟𝑢𝑒) 𝑚𝑎𝑥𝑖𝑚𝑢𝑚 𝑑𝑖𝑠𝑡𝑎𝑛𝑐𝑒 349 350 On this distance metric, lower values indicate a closer correspondence between ping -induced peaks 351 and condition presentation order, which in turn confers stronger evidence for ping -based decoding 352 enhancement. For our statistical evaluation, we used a two-level permutation approach (similar to van 353 Bree et al., 2022). Specifically, we compared the distribution of empirical PODs with PODs calculated 354 across 106 second-level permutations, randomly grabbing from the pool of first-level shuffled decoder 355 time courses. The p-values were defined by the resulting percentile of the empirical POD within the 356 distribution of second-level shuffled PODs (one-sided test, empirical < permuted). 357 For the detection of decoder peaks in this analysis , we detected the maximum peak in the 358 derivative of the cumulative sum of decoding time series. We chose this peak detection method over 359 more standard approaches—such as simply extracting the largest peak from raw decoding series —360 because independent simulations revealed that this algorithm is most powerful at detecting true POD 361 effects, outperforming a range of competing approaches (Supplementary Materials; Section 2). 362 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 23, 2024. ; https://doi.org/10.1101/2024.04.19.590215doi: bioRxiv preprint 11 363

364 Behavioural results 365 As expected in light of our experimental design, participants achieved high memory recognition 366 performance, with scores approaching ceiling across behavioural analyses . First, we found no 367 significant difference in memory performance across participants between the ping (M = 0.980, SE = 368 0.0032) and no ping condition (M = 0.984, SE = 0.006) during the recognition phase (t(27) = -0.745, p 369 = 0.463; Fig. 1B ), suggesting that the decoding analyses that follow are not influenced by absolute 370 inter-condition differences in behaviour. This general near-ceiling performance is also apparent when 371 analysing recognition performance across participants (M = 0.980, SD = 0.015; Fig. 1C) and blocks 372 (M = 0.980, SD = 0.009; Fig. 1D) . Furthermore, participants reported a high rate of remembered to 373 forgotten judgments during the recall phase (M = 0.819; SD = 0.022). The average RT during 374 encoding was 2313 ms for subsequently recognized trials (SD = 1041 ms; n = 2194 trials), and 2472 375 ms for subsequently forgotten trials (SD = 1105 ms; n = 46 trials; Fig. 1E). 376 377 Event-related potentials 378 We observed a robust evoked EEG response after pings (Fig. 2). Specifically, for each of the three 379 stimulus onset asynchrony ( SOA) conditions, w e observed an extended peak of activity across 380 occipitoparietal channels that followed the distribution of ping times for retrieval cue -locked data , 381 peaking approximately 200 to 300 ms after pings. To further confirm that pings successfully evoked a 382 visual response, we applied a ping -locked analysis across all channels and found significantly higher 383 ERP amplitudes after pinged than no-pinged trials in posterior channels (Fig. 2, insets). Together, the 384 ERP analysis suggests pings yielded a strong time-locked response that could putatively interact with 385 ongoing LTM representations. For cue-locked and ping -locked ERPs for each participant, time -386 resolved topographical plots, and for p -values of each channel in Fig . 2 inset topographies, see the 387 Supplementary Materials (Section 1). 388 389 390 amplitude Early ping p < 0.05 Middle ping Late ping cue-lock 00 .5 11 .5 2 time [s] -1 0 1 2 3 4 jump ERP per ping condition .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 23, 2024. ; https://doi.org/10.1101/2024.04.19.590215doi: bioRxiv preprint 12 Figure 2. Ping-induced e vent-related potential . Average evoked response in posterior EEG 391 channels across early (turquoise), middle (blue), and late ping (purple) trials during the recall phase. 392 The inset topographies reveal higher posterior amplitudes following ping trials as contrasted with no-393 ping trials (Monte Carlo permutation test; Bonferroni-corrected). 394 395 Decoding results 396 Stimulus category selection 397 We used a multi -class LDA on no -ping trials (25% of the overall recall trials) to determine which 398 retrieved stimulus category (top, middle, or bottom level) is most robustly decodable when our main 399 experimental pinging manipulation was not present (Fig. 3). We found that none of the three levels 400 displayed significant windows of decodability during our retrieval period of interest from 500 to 2000 401 ms after cue onset (Wilcoxon signed-rank test; p > 0.11 for top; p > 0.25 for middle; p > 0.07 for bot). 402 We proceeded with the top -level, which with its two classes (objects and scenes) afforded simple 403 binary classification with comparatively low variability in decoding performance. Next, during our main 404 analysis, we investigated whether pings enhance the decodability of LTM contents. 405 406 407 Figure 3. Stimulus category selection. Average decoding accuracy across stimulus category levels 408 (top, middle, bottom). Decoding accuracy was quantified relative to the average performance across 409 shuffled decoding results. No significant differences were observed for any level ( Wilcoxon signed 410 rank test, controlled for multiple comparisons using FDR). 411 412 Main analysis 413 For our central analysis, we compared decoder performance between ping and no -ping trials for top-414 level (objects vs scenes) classification, both with the data locked to retrieval cues, and to 415 pings/pseudo-pings (i.e., artificial markers derived from the pool of ping timings; Fig. 4). For the cue -416 locked analysis, we found no windows where decoding was above chance for no-ping trials (two-level 417 Monte Carlo permutation; p > 0.49; Fig. 4A), while the ping trials showed several significant windows 418 of content decodability (p < 0.05; Fig. 4B). To validate our analysis we carried out a direct comparison 419 -0.5 0 0.5 11 . 52 time [s] -0.05 µshuffle 0.05 0.1 accuracy No ping decoder across levels Top Middle Bottom cue-lock jump .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 23, 2024. ; https://doi.org/10.1101/2024.04.19.590215doi: bioRxiv preprint 13 between the ping and no-ping trial decoder, as opposed to contrasting each condition with a shuffled 420 baseline. In this analysis, we found no evidence for a ping-induced decodability enhancement; neither 421 in the cue-locked (Wilcoxon signed-rank test; p > 0.99; Fig. 4C) nor in the (pseudo -)ping-locked data 422 (p > 0.99; Fig. 4D). 423 In light of an important methodological observation, we place more importance on the latter 424 analysis, which directly compares the empirical decoding performance for ping and no-ping conditions 425 without leveraging shuffled results . Specifically, we observed that the standard error of the mean 426 (SEM) of the shuffled distributions varies substantially between ping (μSEM = 0.047) and no-ping (μSEM 427 = 0.028), which we speculated could be explained by trial number differences alone. We inferred that 428 since the ping trial decoder was trained and tested on three times more trials than the no -ping trial 429 decoder, this might naturally shrink SEM values of the shuffled distribution and thereby modulate test 430 statistics. In support of this interpretation, we built a simulation which confirms that an increase in the 431 number of trials (and the number of decoding classes) reduces p-values, but only if there is an effect 432 in the data (Supplementary Materials; Section 3). Therefore, instead of relying on ping -to-shuffle and 433 no-ping-to-shuffle comparisons where power differences might misleadingly lead us to infer a ping-434 related enhancement, we placed most credence in the direct comparison between ping and no -ping 435 trials in which shuffled results are sidestepped (Fig. 4C & Fig. 4D ; see the Supplementary Materials 436 for an extended discussion; Section 3.3). 437 438 439 Figure 4. Main decoder analysis . (A) Cue-locked decoding across no-ping trials compared with a 440 shuffled baseline. (B) Cue-locked decoding across ping trials compared with a shuffled baseline . (C) 441 Direct comparison between on ping and no-ping trials. (D) Same as (C), but with the data time-locked 442 Ping decoderNo ping decoder Ping > no ping decoder Ping > no ping decoder -0.5 0.5 11 . 52 time [s] 0.46 0.48 0.5 0.52 0.54 0.56 0.58 cue-lock jump 0 No ping (HA) Shuffle (H0) area under curvearea under curve Ping No ping -0.5 0 0.5 11 . 52 time [s] 0.46 0.48 0.5 0.52 0.54 0.56 0.58 Ping (HA) Shuffle (H0) p < 0.05 -0.5 0 0.5 1 0.48 0.5 0.52 0.54 0.56 0.48 0.5 0.52 0.54 0.56 -0.5 0 0.5 11 . 5 2 cue-lock jump ping-lockcue-lock jump time [s] time [s] BA DC Ping No ping .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 23, 2024. ; https://doi.org/10.1101/2024.04.19.590215doi: bioRxiv preprint 14 to pings and (artificially marked) pseudo-pings. In (A) and (B) the shaded area represents the 5 th and 443 95th percentile of the distribution of 2nd-level permutations of the shuffled decoder, and in (B) and (C) it 444 represents the SEM of the empirical decoder . In (A) and (B), p -values were derived using two -level 445 Monte Carlo permutations, and in (C) and (D) using Wilcoxon signed -rank test (a ll p -values were 446 corrected using FDR). 447 448 Condition-relative decoding peaks 449 Next, we turn to the presumably more sensitive peak-order analyses. Qualitatively, we observe no 450 ordered structure in decoder peaks when averaging across participants for each SOA pinging 451 condition (Fig. 5A). For a quantitative analysis, we formally compared peak order structure by 452 comparing POD scores for the empirical and shuffled decoder using two-level permutation tests. This 453 analysis confirmed the previous result by revealing no significant evidence for the hypothesis that 454 pings induce systematic differences in the order of decoding peaks (p = 0.357; Fig. 5B). 455 456 457 Figure 5. Condition-relative peak analysis. (A) Decoding results specific to for early (cyan), middle 458 (blue), and late (purple) ping conditions, averaged across participants. (B) Peak order distance scores 459 for the empirical decoder (red line) among a pool of 2 nd-level permutations derived from the shuffled 460 decoder (grey distribution). 461 462

463 In this study, we set out to systematically evaluate visual perturbation, or ping-based stimulation, as a 464

to dynamically enhance the decodability of reactivated neural representations during memory 465 recall. Such an approach could supplement offline analytical approaches by adding further read -out 466 enhancements online at the experiment side. Despite promising results in the WM literature, in this 467 LTM context we found no evidence for a ping -based enhancement across several time -resolved 468 decoding analyses. While pings evoked a strong brain response, they did not detectably boost neural 469 signatures of memory representations in EEG data . We draw this conclusion based on two key 470 results. First, in the main comparison between pinged trials and non -pinged trials, we found no 471 significant decoding difference regardless of whether the data was locked to (pseudo -)pings or 472 retrieval cues. Second, in a more advanced analysis that leverages the constraining information of 473 peak order distance (POD) 0.2 0.4 0.6 0.8 1 2 4 6 8Density (a.u.) Relative decoding peak order p = 0.357 Empirical score Shuffle distribution 0.5 1 1.5 2 time [s] 0.48 0.5 0.52 0.54 0.56area under curve Early ping Middle ping Late ping Decoder results by ping SOA BA .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 23, 2024. ; https://doi.org/10.1101/2024.04.19.590215doi: bioRxiv preprint 15 ping presentation timings during the experiment, we also found no evidence for ping-related decoding 474 increases. 475 There are three overarching explanations for these null results. First, there could be an effect 476 in the data that was left undetected analytically or statistically. Second, there could be an effect that 477 manifests across other experimental contexts, but not with this study’s parameters. Third, there could 478 be no effect in principle, with LTM-based retrieval eluding the enhancement of representational 479 readouts using pings. We consider each option in turn. 480 First, the signal analysis parameter space is high , with variability in parameters across 481 preprocessing and statistical analysis steps potentially altering the results. One important source of 482 variability concerns the implementation of decoding techniques . Namely, we do not rule out that 483 untested decoding methods such as linear approaches beyond LDA or non-linear classifiers would 484 have resulted in performance enhancements induced by pings. More trivially, our analyses could have 485 been optimal, with our key statistical results containing a type-II statistical error. 486 Second, the parameter space on the experimental side is also high. Here, we opted for a 487 word-image association task, which has previously been shown to afford classification -based 488 inferences about memory processing in the brain (Linde-Domingo et al., 2019; Martín -Buro et al., 489 2020; Mirjalili et al., 2021; Kerrén et al., 2022) . However, other LTM tasks might be better suited to 490 reveal ping-based enhancements. Besides the memory task itself, a key set of parameters concerns 491 the presentation of pings. In this study, we chose a high-intensity, short-lasting ping presented with a 492 uniform distribution between 500 and 1500 ms after retrieval cues. This time window was selected 493 based on a review of the timeline of memory reactivation during cued recall, which suggest ed a 494 maximal content reinstatement within this period (Staresina & Wimber, 2019). However, we observed 495 that decoding was highest late within and even after this range, at approximately 1200 – 2000ms after 496 cue (see Fig. 4D). Decoding plateaus that exceed 1500ms have also been observed in recent work 497 that employed a similar task and analysis pipeline (Kerrén et al., 2022). This raises the possibility that 498 the aforementioned 500 to 15 00 ms window is biased to be too early —perhaps because it was 499 estimated based on intracranial EEG research where recordings tend to focus on the hippocampus 500 and other regions that activate early during retrieval (Merkow et al., 2015; Mormann et al., 2005; 501 Staresina et al., 2019). Put differently, it is possible that we did not find significant effects because the 502 signatures of retrieved contents tended to arise robustly only after our ping presentation times . We 503 recommend that future work considers later ping times, potentially informed by maximum decodability 504 periods found in this and other work , or ideally in newly acquired pilot data . Moreover, additional 505 research could explore parameters such as ping duration, intensity, and strength. Furthermore, 506 besides visual pings, a plethora of other perturbational approaches are on stock that could realize the 507 ping’s proposed effects . Also inspired by WM research, stimulation using auditory impulses might 508 offer a multimodal route to improving the readout of LTM contents (Kandemir & Akyürek, 2023) . 509 Furthermore, brain stimulation methods like transcranial magnetic and ultrasound stimulation have the 510 potential to regularize brain activity through the induction of a dynamics -altering magnetic or 511 ultrasound pulse (Moliadze et al., 2003; Mueller et al., 2014). 512 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 23, 2024. ; https://doi.org/10.1101/2024.04.19.590215doi: bioRxiv preprint 16 A third possibility is that none of these factors explain our null results, with ping-based 513 approaches restricting their utility to WM tasks. One specific possibility could be that WM and LTM 514 differ in their mechanisms of action, with separate kinds of neural processes underpinning them. 515 Indeed, classically WM is believed to involve the active maintenance of stimulus-induced information 516 (Fuster & Alexander, 1971; Goldman -Rakic, 1995) , whereas LTM is assumed to be based on a 517 generative reconstruction of past experience based on the activation of silent information-storing 518 engrams (Josselyn & Tonegawa, 2020) . Perhaps the sweep of activity associated with the ping 519 interacts more effectively with functional brain activity maintained continuously from stimulus onset , 520 thus explaining WM-to-LTM differences. Speaking against this interpretation is work that suggests 521 WM representations are encoded in activity -silent networks through short-lasting synaptic changes 522 (Kamiński & Rutishauser, 2020; Masse et al., 2020; Stokes, 2015), which would not be fundamentally 523 different from how LTM works . Contradicting this in turn is a critique which argues that evidence for 524 activity-silent networks in WM tasks could alternatively be explained by LTM processes kicking in 525 (Beukers et al., 2021). Thus, since it is both unclear to what extent the mechanisms of WM and LTM 526 differ and to what extent WM and LTM intertwine in studies where ping -based effects have been 527 demonstrated, we avoid firm interpretations in this part of the possibility space. In summary, although 528 pings unambiguously elicited expected patterns of visual activity (Fig. 2), we failed to find effects on 529 memory decoding, either because they were left undetected in our analysis, because they do not 530 show up in our experimental protocol, or because they do not exist. 531 This study builds on decoding research that investigates the physical basis of memory, 532 leveraging it s findings for a strictly instrumental purpose: the systematic enhancement of LTM 533 readouts. This undertaking is key because the field presently lacks temporally sensitive neuroimaging 534

that enable the consistent and clear readout of memory representations, which is needed to 535 explain how the brain implements memory processes. Furthermore, the analytical challenges, null 536 results, and possible solutions considered in this work could inform practice in fields closely aligned 537 with memory, such as the neuroscience of mental imagery (Dijkstra et al., 2018). 538 To conclude, most efforts to improve memory readouts from electrophysiology data have 539 been restricted to the signal analysis end. Here, we advocate for research that explores online 540 manipulations as memory tasks are unfolding, which has previously shown to complement or 541 synergize with decoding techniques. For long -term memory decoding in particular however, such 542 interventions are scarce, which limits research because memory involves low decodability to begin 543 with. Thus, even if a further carving out of the parameter space does not demonstrate a notable 544 benefit of visual perturbations, future research should creatively explore alternative online methods 545 such as multimodal stimulation and non-invasive brain stimulation. 546 547 Acknowledgments 548 We thank David Rose, Janvi Sidhu, and Jacqueline McDiarmid for their assistance during data 549 acquisition. This work was supported by a Starting Grant from the European Research Council 550 awarded to MW (ERC-2016- StG-715714). 551 552 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 23, 2024. ; https://doi.org/10.1101/2024.04.19.590215doi: bioRxiv preprint 17 Competing interests 553 The authors declare no competing interests. 554 555

556 Barbosa, J., Lozano -Soldevilla, D., & Compte, A. (2021). Pinging the brain with visual impulses 557 reveals electrically active, not activity -silent, working memories. PLoS Biology , 19(10), 558 e3001436. https://doi.org/10.1371/journal.pbio.3001436 559 Benjamini, Y., & Yekutieli, D. (2001). The control of the false discovery rate in multiple testing under 560 dependency. The Annals of Statistics , 29(4), 1165 –1188. 561 https://doi.org/10.1214/aos/1013699998 562 Beukers, A. O., Buschman, T. J., Cohen, J. D., & Norman, K. A. (2021). Is Activity Silent Working 563 Memory Simply Episodic Memory? Trends in Cognitive Sciences , 25(4), 284 –293. 564 https://doi.org/10.1016/j.tics.2021.01.003 565 Brodeur, M. B., Dionne -Dostie, E., Montreuil, T., & Lepage, M. (2010). The Bank of Standardized 566 Stimuli (BOSS), a New Set of 480 Normative Photos of Objects to Be Used as Visual Stimuli 567 in Cognitive Research. PLOS ONE , 5(5), e10773. 568 https://doi.org/10.1371/journal.pone.0010773 569 Cruzat, J., Torralba, M., Ruzzoli, M., Fernández, A., Deco, G., & Soto-Faraco, S. (2021). The phase of 570 Theta oscillations modulates successful memory formation at encoding. Neuropsychologia, 571 154, 107775. https://doi.org/10.1016/j.neuropsychologia.2021.107775 572 Danker, J. F., & Anderson, J. R. (2010). The ghosts of brain states past: Remembering reactivates 573 the brain regions engaged during encoding. Psychological Bulletin , 136(1), 87 –102. 574 https://doi.org/10.1037/a0017937 575 Deuker, L., Olligs, J., Fell, J., Kranz, T. A., Mormann, F., Montag, C., Reuter, M., Elger, C. E., & 576 Axmacher, N. (2013). Memory Consolidation by Replay of Stimulus -Specific Neural Activity. 577 Journal of Neuroscience , 33(49), 19373 –19383. https://doi.org/10.1523/JNEUROSCI.0414-578 13.2013 579 Dijkstra, N., Mostert, P., Lange, F. P. de, Bosch, S., & van Gerven, M. A. (2018). Differential temporal 580 dynamics during visual imagery and perception. eLife, 7, e33904. 581 https://doi.org/10.7554/eLife.33904 582 Duncan, D. H., van Moorselaar, D., & Theeuwes, J. (2023). Pinging the brain to reveal the hidden 583 attentional priority map using encephalography. Nature Communications , 14(1), Article 1. 584 https://doi.org/10.1038/s41467-023-40405-8 585 Favila, S. E., Kuhl, B. A., & Winawer, J. (2022). Perception and memory have distinct spatial tuning 586 properties in human visual cortex. Nature Communications , 13(1), 5864. 587 https://doi.org/10.1038/s41467-022-33161-8 588 Favila, S. E., Lee, H., & Kuhl, B. A. (2020). Transforming the Concept of Memory Reactivation. 589 Trends in Neurosciences, 43(12), 939–950. https://doi.org/10.1016/j.tins.2020.09.006 590 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 23, 2024. ; https://doi.org/10.1101/2024.04.19.590215doi: bioRxiv preprint 18 Fritch, H. A., MacEvoy, S. P., Thakral, P. P., Jeye, B. M., Ross, R. S., & Slotnick, S. D. (2020). The 591 anterior hippocampus is associated with spatial memory encoding. Brain Research, 1732, 592 146696. https://doi.org/10.1016/j.brainres.2020.146696 593 Fuster, J. M., & Alexander, G. E. (1971). Neuron Activity Related to Short -Term Memory. Science, 594 173(3997), 652–654. https://doi.org/10.1126/science.173.3997.652 595 Goldman-Rakic, P. S. (1995). Cellular basis of working memory. Neuron, 14(3), 477 –485. 596 https://doi.org/10.1016/0896-6273(95)90304-6 597 Grootswagers, T., Wardle, S. G., & Carlson, T. A. (2017). Decoding Dynamic Brain Patterns from 598 Evoked Responses: A Tutorial on Multivariate Pattern Analysis Applied to Time Series 599 Neuroimaging Data. Journal of Cognitive Neuroscience , 29(4), 677 –697. 600 https://doi.org/10.1162/jocn_a_01068 601 Haque, R. U., Wittig, J. H., Damera, S. R., Inati, S. K., & Zaghloul, K. A. (2015). Cortical Low -602 Frequency Power and Progressive Phase Synchrony Precede Successful Memory Encoding. 603 Journal of Neuroscience , 35(40), 13577 –13586. https://doi.org/10.1523/JNEUROSCI.0687-604 15.2015 605 Haxby, J. V., Connolly, A. C., & Guntupalli, J. S. (2014). Decoding Neural Representational Spaces 606 Using Multivariate Pattern Analysis. Annual Review of Neuroscience , 37(1), 435 –456. 607 https://doi.org/10.1146/annurev-neuro-062012-170325 608 Josselyn, S. A., & Tonegawa, S. (2020). Memory engrams: Recalling the past and imagining the 609 future. Science, 367(6473), eaaw4325. https://doi.org/10.1126/science.aaw4325 610 Kamiński, J., & Rutishauser, U. (2020). Between persistently active and activity -silent frameworks: 611 Novel vistas on the cellular basis of working memory. Annals of the New York Academy of 612 Sciences, 1464(1), 64–75. https://doi.org/10.1111/nyas.14213 613 Kandemir, G., & Akyürek, E. G. (2023). Impulse perturbation reveals cross -modal access to sensory 614 working memory through learned associations. NeuroImage, 274, 120156. 615 https://doi.org/10.1016/j.neuroimage.2023.120156 616 Kandemir, G., Wilhelm, S. A., Axmacher, N., & Akyürek, E. G. (2023). Maintenance of colour 617 memoranda in activity-quiescent working memory states: Evidence from impulse perturbation 618 (p. 2023.07.03.547526). bioRxiv. https://doi.org/10.1101/2023.07.03.547526 619 Kayser, J., & Tenke, C. E. (2015). On the benefits of using surface Laplacian (Current Source 620 Density) methodology in electrophysiology. International Journal of Psychophysiology : 621 Official Journal of the International Organization of Psychophysiology , 97(3), 171 –173. 622 https://doi.org/10.1016/j.ijpsycho.2015.06.001 623 Kerrén, C., van Bree, S., Griffiths, B. J., & Wimber, M. (2022). Phase separation of competing 624 memories along the human hippocampal theta rhythm. eLife, 11, e80633. 625 https://doi.org/10.7554/eLife.80633 626 Kragel, J. E., Ezzyat, Y., Sperling, M. R., Gorniak, R., Worrell, G. A., Berry, B. M., Inman, C., Lin, J. -627 J., Davis, K. A., Das, S. R., Stein, J. M., Jobst, B. C., Zaghloul, K. A., Sheth, S. A., Rizzuto, D. 628 S., & Kahana, M. J. (2017). Similar patterns of neural activity predict memory function during 629 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 23, 2024. ; https://doi.org/10.1101/2024.04.19.590215doi: bioRxiv preprint 19 encoding and retrieval. NeuroImage, 155, 60 –71. 630 https://doi.org/10.1016/j.neuroimage.2017.03.042 631 Kuhl, B. A., Rissman, J., & Wagner, A. D. (2012). Multi-voxel patterns of visual category 632 representation during episodic encoding are predictive of subsequent memory. 633 Neuropsychologia, 50(4), 458–469. https://doi.org/10.1016/j.neuropsychologia.2011.09.002 634 Linde-Domingo, J., Treder, M. S., Kerrén, C., & Wimber, M. (2019). Evidence that neural information 635 flow is reversed between object perception and object reconstruction from memory. Nature 636 Communications, 10(1), 179. https://doi.org/10.1038/s41467-018-08080-2 637 Madore, K. P., & Wagner, A. D. (2022). Readiness to remember: Predicting variability in episodic 638 memory. Trends in Cognitive Sciences , 26(8), 707 –723. 639 https://doi.org/10.1016/j.tics.2022.05.006 640 Maguire, E. A. (2014). Memory consolidation in humans: New evidence and opportunities. 641 Experimental Physiology, 99(3), 471–486. https://doi.org/10.1113/expphysiol.2013.072157 642 Martín-Buro, M. C., Wimber, M., Henson, R. N., & Staresina, B. P. (2020). Alpha Rhythms Reveal 643 When and Where Item and Associative Memories Are Retrieved. The Journal of 644 Neuroscience: The Official Journal of the Society for Neuroscience , 40(12), 2510 –2518. 645 https://doi.org/10.1523/JNEUROSCI.1982-19.2020 646 Masse, N. Y., Rosen, M. C., & Freedman, D. J. (2020). Reevaluating the Role of Persistent Neural 647 Activity in Short -Term Memory. Trends in Cognitive Sciences , 24(3), 242 –258. 648 https://doi.org/10.1016/j.tics.2019.12.014 649 Merkow, M. B., Burke, J. F., & Kahana, M. J. (2015). The human hippocampus contributes to both the 650 recollection and familiarity components of recognition memory. Proceedings of the National 651 Academy of Sciences, 112(46), 14378–14383. https://doi.org/10.1073/pnas.1513145112 652 Mirjalili, S., Powell, P., Strunk, J., James, T., & Duarte, A. (2021). Context Memory Encoding and 653 Retrieval Temporal Dynamics are Modulated by Attention across the Adult Lifespan. eNeuro, 654 8(1), ENEURO.0387-20.2020. https://doi.org/10.1523/ENEURO.0387-20.2020 655 Moliadze, V., Zhao, Y., Eysel, U., & Funke, K. (2003). Effect of transcranial magnetic stimulation on 656 single-unit activity in the cat primary visual cortex. The Journal of Physiology, 553(Pt 2), 665–657 679. https://doi.org/10.1113/jphysiol.2003.050153 658 Mormann, F., Fell, J., Axmacher, N., Weber, B., Lehnertz, K., Elger, C. E., & Fernández, G. (2005). 659 Phase/amplitude reset and theta –gamma interaction in the human medial temporal lobe 660 during a continuous word recognition memory task. Hippocampus, 15(7), 890 –900. 661 https://doi.org/10.1002/hipo.20117 662 Mueller, J., Legon, W., Opitz, A., Sato, T. F., & Tyler, W. J. (2014). Transcranial Focused Ultrasound 663 Modulates Intrinsic and Evoked EEG Dynamics. Brain Stimulation , 7(6), 900 –908. 664 https://doi.org/10.1016/j.brs.2014.08.008 665 Oostenveld, R., Fries, P., Maris, E., & Schoffelen, J. -M. (2011). FieldTrip: Open source software for 666 advanced analysis of MEG, EEG, and invasive electrophysiological data. Computational 667 Intelligence and Neuroscience, 2011, 156869. https://doi.org/10.1155/2011/156869 668 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 23, 2024. ; https://doi.org/10.1101/2024.04.19.590215doi: bioRxiv preprint 20 Pearson, J., Naselaris, T., Holmes, E. A., & Kosslyn, S. M. (2015). Mental Imagery: Functional 669 Mechanisms and Clinical Applications. Trends in Cognitive Sciences , 19(10), 590 –602. 670 https://doi.org/10.1016/j.tics.2015.08.003 671 Peirce, J., Gray, J. R., Simpson, S., MacAskill, M., Höchenberger, R., Sogo, H., Kastman, E., & 672 Lindeløv, J. K. (2019). PsychoPy2: Experiments in behavior made easy. Behavior Research 673 Methods, 51(1), 195–203. https://doi.org/10.3758/s13428-018-01193-y 674 Rissman, J., & Wagner, A. D. (2012). Distributed representations in memory: Insights from functional 675 brain imaging. Annual Review of Psychology , 63, 101–128. https://doi.org/10.1146/annurev-676 psych-120710-100344 677 Rizzuto, D. S., Madsen, J. R., Bromfield, E. B., Schulze -Bonhage, A., Seelig, D., Aschenbrenner -678 Scheibe, R., & Kahana, M. J. (2003). Reset of human neocortical oscillations during a working 679 memory task. Proceedings of the National Academy of Sciences , 100(13), 7931 –7936. 680 https://doi.org/10.1073/pnas.0732061100 681 Schreiner, T., Petzka, M., Staudigl, T., & Staresina, B. P. (2021). Endogenous memory reactivation 682 during sleep in humans is clocked by slow oscillation -spindle complexes. Nature 683 Communications, 12(1), Article 1. https://doi.org/10.1038/s41467-021-23520-2 684 Staresina, B. P., Reber, T. P., Niediek, J., Boström, J., Elger, C. E., & Mormann, F. (2019). 685 Recollection in the human hippocampal -entorhinal cell circuitry. Nature Communications , 686 10(1), 1503. https://doi.org/10.1038/s41467-019-09558-3 687 Staresina, B. P., & Wimber, M. (2019). A Neural Chronometry of Memory Recall. Trends in Cognitive 688 Sciences, 23(12), 1071–1085. https://doi.org/10.1016/j.tics.2019.09.011 689 Stokes, M. G. (2015). “Activity -silent” working memory in prefrontal cortex: A dynamic coding 690 framework. Trends in Cognitive Sciences , 19(7), 394 –405. 691 https://doi.org/10.1016/j.tics.2015.05.004 692 Ten Oever, S., De Weerd, P., & Sack, A. T. (2020). Phase -dependent amplification of working 693 memory content and performance. Nature Communications , 11(1), 1832. 694 https://doi.org/10.1038/s41467-020-15629-7 695 Ter Wal, M., Linde-Domingo, J., Lifanov, J., Roux, F., Kolibius, L. D., Gollwitzer, S., Lang, J., Hamer, 696 H., Rollings, D., Sawlani, V., Chelvarajah, R., Staresina, B., Hanslmayr, S., & Wimber, M. 697 (2021). Theta rhythmicity governs human behavior and hippocampal signals during memory -698 dependent tasks. Nature Communications, 12(1), 7048. https://doi.org/10.1038/s41467-021-699 27323-3 700 Treder, M. S. (2020). MVPA -Light: A Classification and Regression Toolbox for Multi -Dimensional 701 Data. In Frontiers in Neuroscience (Vol. 14). 702 https://www.frontiersin.org/article/10.3389/fnins.2020.00289 703 van Bree, S., Melcón, M., Kolibius, L. D., Kerrén, C., Wimber, M., & Hanslmayr, S. (2022). The brain 704 time toolbox, a software library to retune electrophysiology data to brain dynamics. Nature 705 Human Behaviour, 6(10), 1430–1439. https://doi.org/10.1038/s41562-022-01386-8 706 Wilcoxon, F. (1945). Individual Comparisons by Ranking Methods. Biometrics Bulletin, 1(6), 80–83. 707 https://doi.org/10.2307/3001968 708 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 23, 2024. ; https://doi.org/10.1101/2024.04.19.590215doi: bioRxiv preprint 21 Wolff, M. J., Ding, J., Myers, N. E., & Stokes, M. G. (2015). Revealing hidden states in visual working 709 memory using electroencephalography. Frontiers in Systems Neuroscience , 9. 710 https://doi.org/10.3389/fnsys.2015.00123 711 Wolff, M. J., Jochim, J., Akyürek, E. G., Buschman, T. J., & Stokes, M. G. (2020). Drifting codes within 712 a stable coding scheme for working memory. PLOS Biology , 18(3), e3000625. 713 https://doi.org/10.1371/journal.pbio.3000625 714 Wolff, M. J., Jochim, J., Akyürek, E. G., & Stokes, M. G. (2017). Dynamic hidden states underlying 715 working-memory-guided behavior. Nature Neuroscience , 20(6), 864 –871. 716 https://doi.org/10.1038/nn.4546 717 Xiao, J., Hays, J., Ehinger, K. A., Oliva, A., & Torralba, A. (2010). SUN database: Large -scale scene 718 recognition from abbey to zoo. 2010 IEEE Computer Society Conference on Computer Vision 719 and Pattern Recognition, 3485–3492. https://doi.org/10.1109/CVPR.2010.5539970 720 Xie, J., & Qiu, Z. (2007). The effect of imbalanced data sets on LDA: A theoretical and empirical 721 analysis. Pattern Recognition, 40(2), 557–562. https://doi.org/10.1016/j.patcog.2006.01.009 722 Xue, G. (2018). The Neural Representations Underlying Human Episodic Memory. Trends in 723 Cognitive Sciences, 22(6), 544–561. https://doi.org/10.1016/j.tics.2018.03.004 724 Yang, C., He, X., & Cai, Y. (2023). Reactivating and reorganizing activity-silent working memory: Two 725 distinct mechanisms underlying pinging the brain (p. 2023.07.16.549254). bioRxiv. 726 https://doi.org/10.1101/2023.07.16.549254 727 728 729 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 23, 2024. ; https://doi.org/10.1101/2024.04.19.590215doi: bioRxiv preprint 22 Supplementary Materials 730 1. Event-related potential 731 732 733 Supplementary Figure 1. Retrieval cue-locked ERP . The purple trace reflects the average cue -734 locked response for each participant across posterior EEG channels. The grey horizontal line 735 represents cue onset. For more details, see the Methods section in the main text. The amplitude on 736 the y-axis is in arbitrary units. 737 738 739 740 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 23, 2024. ; https://doi.org/10.1101/2024.04.19.590215doi: bioRxiv preprint 23 Supplementary Figure 2. Retrieval ping-locked ERP. The purple trace reflects the average ping-741 locked response for each participant across posterior EEG channels. The grey horizontal line 742 represents ping onset. For more details, see the Methods section in the main text. The amplitude on 743 the y-axis is in arbitrary units. 744 745 746 747 748 Supplementary Figure 3. Retrieval cue-locked topographies. These topographical plots represent 749 the average cue -locked activity across participants. The colours represent the difference in EEG 750 activity before and after cue onset in arbitrary units (red colours represent activity post > activitypre and 751 vice versa for blue colours). No statistical analysis was carried out for these topographical contrasts. 752 For more details, see the Methods section in the main text. 753 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 23, 2024. ; https://doi.org/10.1101/2024.04.19.590215doi: bioRxiv preprint 24 754 755 Supplementary Figure 4. Retrieval ping -locked topographies (ping vs. no ping trials) . These 756 topographical plots represent the average cue -locked activity across participants. The colours 757 represent the difference in EEG activity between ping and no-ping (red colours represent activityping > 758 activityno ping and vice versa for blue colours). No statistical analysis was carried out for these 759 topographical contrasts. For more details, see the Methods section in the main text. 760 761 762 Channel Early ping (p-val) Middle ping (p-val) Late ping (p-val) Fp1 0.032 0.616 0.246 Fpz 0.089 0.079 0.011 Fp2 0.042 0.537 0.115 AF8 0.119 0.422 0.318 AF7 0.014 0.272 0.954 AF3 0.439 0.23 0.123 AF4 0.712 0.541 0.014 F7 0.002 0.346 0.358 F5 0.068 0.439 0.33 F3 0.119 0.477 0.693 F1 0.597 0.662 0.119 Fz 0.053 0.551 0.003 F2 0.013 0.473 0 F4 0.341 0.939 0.049 F6 0.427 0.559 0.707 F8 0.131 0.826 0.825 FT8 0.001 0.097 0.049 FC6 0.177 0.142 0.78 FC4 0.962 0.176 0.881 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 23, 2024. ; https://doi.org/10.1101/2024.04.19.590215doi: bioRxiv preprint 25 FC2 0.245 0.276 0.09 FC1 0.969 0.503 0.881 FC3 0.176 0.279 0.28 FC5 0.011 0.083 0.112 FT7 0.002 0.298 0.127 T7 0.004 0.047 0.043 C5 0.013 0.027 0.051 C3 0.002 0.022 0.01 C1 0.148 0.049 0.04 Cz 0.144 0.155 0.114 C2 0.305 0.182 0.104 C4 0.02 0.004 0.014 C6 0.003 0 0.003 T8 0.002 0.002 0.003 TP10 0.002 0 0 TP8 0 0 0 CP6 0 0 0 CP4 0.001 0 0.001 CP2 0.006 0.001 0.001 CPz 0.019 0.006 0.01 CP1 0.272 0.002 0.003 CP3 0 0 0.001 CP5 0.002 0.001 0.001 TP7 0.002 0.008 0.001 TP9 0 0 0 P7 0 0 0 P5 0 0 0 P3 0 0 0 P1 0 0 0 Pz 0.001 0 0 P2 0.001 0 0 P4 0 0 0 P6 0 0 0 P8 0 0 0 PO8 0 0 0 PO4 0 0 0 POz 0 0 0 PO3 0 0 0 PO7 0 0 0 O1 0 0 0 Oz 0.001 0 0 O2 0 0 0 763 Supplementary Table 1. P-values associated with inset topographies in main text Fig. 2; rounded to 764 three decimal points. 765 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 23, 2024. ; https://doi.org/10.1101/2024.04.19.590215doi: bioRxiv preprint 26 2. Peak order analysis simulation 766 2.1 Time series simulation 767 We used MATLAB (the MathWorks) to generate time series with two components: (1) a peak at a 768 fixed time point (1000 ms), and (2) autocorrelated noise generated using a random walk procedure. 769 We matched several characteristics of the simulated time series to our empirical decoding data, 770 including the analysis period (500 to 2000 ms) , sampling rate (50 Hz) , and the number of (virtual) 771 participants (N = 29) . The signal -to-noise (SNR) ratio of the simulation was set to 1.15 , qualitatively 772 matching peaks observed in the empirical data. We found that varying the SNR does not significantly 773 alter the results. We generated 1000 trials per participant, resulting in 29000 trials in total. 774 775 2.2 Analysis 776 We included a smoothing parameter that implemented one of four smoothing methods : no filter, a 777 Gaussian filter, a Savitzky-Golay filter, and a median filter. We also included a window size for 778 smoothing, set to 10 samples for our main analysis. We compared the performance of eight peak 779 detection methods, evaluating each of them based on the absolute distance between estimated peaks 780 and true peaks—amounting to a simplified version of the peak order distance score described under 781 condition-relative decoding peaks in the main text . The winning method was locked in for our 782 empirical analysis. We tested eight peak detection methods: 783 (1) Low-pass approach, where the maximum peak was computed after a low -pass filter was 784 applied to the time series. 785 (2) Maximum value approach, which simply computed the maximum value per time series 786 regardless of whether the surrounding data was peak-like. 787 (3) Cumulative sum approach, which computed the maximum peak in the derivative of the 788 cumulative sum of the data. 789 (4) Cumulative integral approach, which computed the maximum peak in the cumulative integral 790 of the data via the trapezoidal method. 791 (5) Integral cumulative sum approach, which worked as the previous method but which operates 792 over the cumulative sum rather than raw time series. 793 (6) Wavelet transform-based method, which finds the maximum peak in a wavelet decomposed 794 version of the data. 795 (7) Hilbert transform-based method, which find the maximum peak in the amplitude fluctuations in 796 the envelope of the time series. 797 (8) Cross-correlation method, which finds the time lag with a maximal correlation between the 798 signal and iteratively shifted versions of itself. 799 800 2.3 Results 801 We found that approach 5—the i ntegral cumulative sum approach —reliably achieves low absolute 802 distance errors across parameters (Supplementary Figure 5). These results were generally 803 unchanged across adjustments of the parameters (to evaluate this, we refer to the code published 804 with this manuscript). Thus, we used approach 5 in our main peak order detection analysis. 805 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 23, 2024. ; https://doi.org/10.1101/2024.04.19.590215doi: bioRxiv preprint 27 806 Supplementary Figure 6. In simulated time series, the integral cumulative sum approach works best 807 for detecting a peak in noisy time series. The red circle indicates the best -performing method, and 808 yellow the second best-performing method. Errors were computed based on the absolute distance in 809 milliseconds (ms) between estimated and true peak location. 810 811 3. Class and trial number decoding simulation 812 We speculated based on a qualitative inspection of the empirical decoding results that the number of 813 trials (Ntrials) and classes (Nclasses) reduces the statistical significance of decoding results. We 814 evaluated this intuition by demonstrating using simulations that these two parameters do indeed 815 influence the variance of shuffled and empirical results, which in turn affects p-values but only if there 816 is a true effect in the data. 817 818 3.1 Time series simulation 819 Using MATLAB, w e generated one ground truth vector of class labels which represented the true 820 class structure in the simulated data. This vector contained a random sequence of integers randomly 821 grabbed between the interval 1 and Nclasses. For example, with 16 classes, the ground truth pattern 822 might have contained a sequence of [2,7,15,4,13,17] and with 2 classes a sequence of [2,2,1,2,1,2]. 823 Then, to simulate shuffled decoding results, we generated a distribution of random sequences 824 of integers identical to the ground truth procedure, but with newly generated random integers. These 825 random sequences represented shuffled decoding results and were scored based on their average 826 element-wise correspondence to the ground truth pattern —which is how decoding accuracy is 827 normally computed. For example, if the permuted vector is [2,1,2,2,1,1] and the true sequence is 828 [2,2,1,2,1,2], the accuracy would be 50% because half of the class labels correspond to the true 829 structure. Trivially, with increasing repetitions the shuffled distribution will approach chance level 830 predictions of the ground truth pattern (i.e., the expected value is exactly at 1/Nclasses). 831 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 23, 2024. ; https://doi.org/10.1101/2024.04.19.590215doi: bioRxiv preprint 28 Finally, to simulate empirical decoding results, we again generated a distribution of random 832 integers identical to the procedure for shuffled and ground truth decoding results. However, for these 833 data we manually injected between 0% and 60% of the ground truth pattern into the otherwise 834 random vector, effectively modulating decoding accuracy. With 0% of the ground truth injected, there 835 is no statistically detectable difference in accuracy between empirical and shuffled decoding results, 836 because the vectors are equally random. With 60%, the encoding results are substantially more 837 accurate than shuffled results, yielding above chance decoding accuracy. 838 We simplified our simulation by operationalizing the variable Ntrials as the number of elements 839 in the vector , allowing us to efficiently investigate how the number of observations influences 840 statistical tests. We also compared Nclasses = 2 and Nclasses = 16, which respectively match the number 841 of classes for top- and bottom-level category decoding in our main experiment. Both Ntrials and Nclasses 842 were independently manipulated in a 2 ∗ 2 factorial design, allowing us to evaluate the contribution of 843 each variable toward statistical outcomes (as a function of effect size). 844 845 3.2 Results 846 First, with respect to N classes, we found that increasing the number of classes reduces the spread of 847 both shuffled and empirical decoding results (Supplementary Figure 7; columns). This happens both if 848 there is no true effect in the empirical data, and when a significant proportion of the ground truth is 849 inserted into the empirical data . Second, we found that Ntrials similarly reduces the variance of both 850 shuffled and decoding results, both across low and high N classes (Supplementary Figure 7; top and 851 bottom half). Thus, we conclude that both factors modulate the likelihood of finding a significant 852 difference between empirical and shuffled results, but only if there is a true effect in the data. Indeed, 853 as we can glean from the results based on non -existent effects, the distributions of empirical and 854 shuffled will overlap regardless of N trials or Nclasses (Supplementary Figure 7; left half). In contrast, if 855 there is an effect (60% injected ground truth), both N trials and N classes independently increase the 856 distributional distance between empirical and shuffled accuracy values. 857 858 3.3 Discussion 859 We found that N trials and N classes independently reduce the variance of accuracy results, which will 860 affect statistical tests between empirical and shuffled distributions but only if there is an effect in the 861 data. As suggested in the main text, these findings suggest that statistical analyses that depend on 862 variance comparisons between empirical and shuffled distributions should be interpreted with care if it 863 is done across conditions with varying Ntrials and Nclasses. With regard to our main analysis for example, 864 the fact that the decoder based on pinged trials yields more significant decodability compared to the 865 decoder based on no-pinged trials should be interpreted with caution because there are differences in 866 Ntrials between the two conditions that could partially or fully explain this effect. More generally, we 867 found that the condition with more trials or more classes is by default more likely to yield significant p-868 values—but only if a true effect exist. 869 870 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 23, 2024. ; https://doi.org/10.1101/2024.04.19.590215doi: bioRxiv preprint 29 871 Supplementary Figure 7. The effects of class and trial number on decoding accuracy. Both the 872 number of classes (columns) and trials (top vs. bottom half) influences the distance between shuffled 873 and empirical distributions—but only if there is an effect in the data (left vs. right half). 874 875 These findings may be a manifestation of the classical notion of statistical power in statistical 876 analysis but within the less intuitive context of decoding accuracy . Our interpretation then is not that 877 Ntrials and N classes must necessarily be equal between conditions for a statistical comparison to be 878 meaningful. Rather, we wanted to err on the side of caution and ensure that analyses where power 879 differences could possibly explain condition differences (e.g., Fig. 3 and Fig. 4A and 4B in the main 880 text) do not inform subsequent analyses and scientific interpretations by themselves . Instead, we 881 supplemented each of the implicated analyses with additional rationale (in the case of Fig. 3) or 882 analyses that do not involve empirical -to-shuffle decoding comparisons. Indeed, Fig. 4C and Fig. 4D 883 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 23, 2024. ; https://doi.org/10.1101/2024.04.19.590215doi: bioRxiv preprint 30 involve direct comparisons between empirical and shuffled distributions , sidestepping the issue 884 altogether. 885 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 23, 2024. ; https://doi.org/10.1101/2024.04.19.590215doi: bioRxiv preprint