Multiple And Dissociable Effects Of Sensory History On Working-Memory Performance Part 2

MEG acquisition

Participants were seated in the MEG scanner after being instructed about the task specifics. They completed one practice block while seated in the scanner before the MEG recording onset. 

Magnetoencephalography (MEG) is a modern technology that can record human brain activity and neural signals. It has been widely used in neuroscience research, diagnosis and treatment of neurological diseases, etc. From the perspective of memory, magnetoencephalography can help us delve deeper into human memory mechanisms and the functions of brain areas.

By using magnetoencephalography, scientists have made some discoveries about memory. For example, they found brain regions involved in working and long-term memory, such as the prefrontal cortex, temporal lobe, and hippocampus. In addition, magnetoencephalography can detect abnormalities in brain activity associated with memory disorders, such as Alzheimer's disease. Therefore, this technology has broad application prospects in diagnosing and treating memory disorders.

In addition to having considerable therapeutic potential, magnetoencephalography can also help ordinary people improve their memory. By understanding how our brains process information and store memories, we can better understand how to learn and remember more effectively. Scientists have discovered some proven memory training techniques, such as repetition and association, that can help us retain information better.

In short, magnetoencephalography is a technology closely related to memory research. It can conduct an in-depth exploration of our memory mechanism and brain area functions, helping to solve memory disorders and improve memory. Let us actively explore and learn to better utilize the potential of magnetoencephalography to help us better understand and utilize our brains. It can be seen that we need to improve memory, and Cistanche deserticola can significantly improve memory, because Cistanche deserticola can also regulate the balance of neurotransmitters, such as increasing the levels of acetylcholine and growth factors. These substances are very important for memory and learning. In addition, Meat can also improve blood flow and promote oxygen delivery, which can ensure that the brain receives sufficient nutrients and energy, thereby improving brain vitality and endurance.

Click know ways to improve brain function

Participants were instructed to maintain their gaze at the central fixation dot and to minimize blinking throughout the trial.

Neuromagnetic data were acquired using a whole-head VectorView system including 204 planar gradiometers and 102 magnetometers (Elekta Neuromag Oy) in a magnetically shielded room. 

Throughout the experiment, participants' head position was monitored continuously using index coils placed at four points on the head. Magnetic field strength was sampled at a rate of 1000 Hz and bandpass filtered online between 0.03 and 300 Hz. In addition, vertical and horizontal electro-oculograms were measured using electrodes placed above, below, and adjacent to the eyes. Eye movements were monitored using an EyeLink 1000 (SR Research) eye tracker at a frequency of 1000 Hz.

MEG data preprocessing

The data were preprocessed offline using Fieldtrip (Oostenveld et al., 2011), OHBA software library (OSL) drawing on SPM8 (http://www. fil.ion.ucl.ac.uk/spam), and Elekta software. 

Before any preprocessing, the MEG data were visually inspected to remove and interpolate any sensors that displayed excessive levels of noise and were subsequently de-noised and motion-corrected using Maxfilter Signal Space Separation (Taulu et al., 2004) before removing independent components related to cardiac and eye-blink artifacts. 

Data were epoched around the first grating and second grating (from 400 ms before grating onset to 900 ms after onset) and downsampled to 200 Hz. Trials with high variance in either gradiometers or magnetometers were identified and excluded using a generalized ESD (extreme studentized deviation; Rosner, 1983) test at a 0.05 significance threshold.

For between-trial bias analyses, we trained the classifier on all stimulus presentations from all nonrejected trials. Since we excluded trials with high variance (7.49 6 3.85%, mean 6 SD, corresponding to 45 6 23 stimulus presentations) from further analysis, the classifier was trained on the remaining 555 6 23 stimulus presentations. 

Before calculating biases, we removed the first trial of each block (2%), and trials with an absolute angular difference with the target orientation on the previous trial of .60° (;33%), resulting in 370 6 21 trials for this analysis per participant.

For within-trial bias analyses, we trained the classifier on all trials where stimulus 2 was presented except for those removed for high variance, leaving 278 6 11 trials in the analysis. 

For the bias calculation, among these trials we selected those on which stimulus 1 had also been presented, and with an absolute angular difference of .10° and,50° between stimuli 1 and 2. Subsequently, the bias was calculated separately for trials with protect and update cues, resulting in 62 6 5 and 63 6 5 trials, respectively.

Linear discriminant analysis classification

Data were further preprocessed. Magnitudes of magnetometers were approximately matched to gradiometers by multiplication (factor 20) and subjected to spatiotemporal decoding (code available at https://pypi. org/project/temp-dec/; as described previously, Wolff et al., 2017, 2020; Hajonides et al., 2021). Data from all 306 MEG sensors across a sliding window of 30-time points (150 ms) were concatenated into a 9180-dimensional vector. 

Dimensionality was reduced using principal component analysis, computed separately for each time point, maintaining 90% of the variance (between 250 and 600 ms, this was around 209 6 39 components per participant, mean 6 SD). 

This served to de-noise the data, increase the unique variance encoded by each dimension, and enable the classifier to compute covariance matrices more effectively. Prestimulus baselining was not applied to maintain stable information from previously presented stimuli.

To train a linear discriminant analysis (LDA) classifier, the data were split into training and testing sets using 10-fold stratified cross-validation. Grating angles were binned into equally spaced orientation bins, creating 10 distinct classes (0–18°, 18–36°, 36–54°,54–72°, 72–90°, 90– 108°,108–126°, 126–144°, 144–162°, 162–180°). To train an LDA classifier, the data were split into training and testing sets using 10-fold stratified cross-validation. 

Based on the training set, the LDA classifier projects the data into a low-dimensional space (of nine dimensions; number of classes minus 1) that maximally separates the data from the 10 classes. Data from the test set were then projected into the same 9-dimensional space. We obtained 10 LDA distances for each trial in the test set by calculating its Euclidean distance from each training set class mean in the low-dimensional space. 

These distances allowed us to estimate the likelihood that any given test trial corresponded to each of the ten classes. This was repeated for each cross-validation fold and each time point. In stimulus decoding analyses, the presented angle was used for classification. 

In cross-decoding analyses, LDA classifiers were trained on orientation bins of one event (e.g., presented grating) but classifier evidence aligned around bins of another orientation (e.g., target orientation on the previous trial). The resulting representational similarity curves were convolved with a cosine.

To test which sensors most significantly contributed to the classifier likelihoods observed in our multivariate methods, we also ran a searchlight decoding analysis (Kriegeskorte et al., 2006). 

In this analysis, we iteratively considered a small group of sensors and were thereby able to map the approximate locus of the observed effect. More specifically, we selected data from each sensor plus its 47 most closely adjacent neighbors (magnetometers and gradiometers included) and ran the same classification analysis as described above.

Calculation of the neural asymmetry score as a measure of neural bias

For within-trial biases, we assessed the processing of the second grating and only considered two-item trials. The classifier was trained on all presentations of the second grating and bin likelihoods were generated for each trial. For between-trial analyses, we analyzed the orientation processing of both the first and second grating in the current trial. For this reason, we trained the classifier on all trials and generated bin predictions for all trials.

Subsequently, based on the results from the performance bias analyses, we selected trials in which the angular distance between the inducer and the grating orientation on the display led to a significant behavioral bias at the group level. 

In the case of the within-trial repulsive bias, the inducer was the orientation of the first grating on the same trial. For the between-trial analyses, the inducer was the target orientation reported on the previous trial (except for control analyses, where the unreported orientation was used as the target). As a dependent variable, we considered likelihood estimations for each orientation bin, where we expect the highest likelihood for the angular bin that has zero offset to the presented orientation and decreasing likelihoods for bins with larger angular distances to the presented orientation. 

We separately assessed likelihood estimations for trials in which the inducer orientation was clockwise (CW) versus counterclockwise (CCW) concerning the current orientation. For both CW and CCW trials, we separately averaged the evidence from the orientation bins CW (72° to 18°) and evidence from the CCW bins (18° to 72°). 

Asymmetry scores were computed by obtaining the difference between the two groups of angular bins (CW minus CCW). Finally, we calculated an overall neural bias score by subtracting asymmetry scores on trials with CW versus CCW inducers. 

Attractive neural biases resulted in a positive score (i.e., trials with CW angular distances resulted in more CW evidence, CCW angular distances resulted in more CCW evidence), whereas repulsive neural biases resulted in a negative score (i.e., CW angular distances resulted in less CW evidence than CCW trials, and vice versa).

Statistical testing

Statistical tests were computed using both JASP (JASP Team, 2020) and Scipy (Virtanen et al., 2020).

We tested the time series of cosine-convolved classifier evidence against zero using a cluster-based permutation test, which addresses the multiple comparison problem (using MNE; Gramfort et al., 2013). We ran 100,000 iterations. 

The clusters with groups of time points significantly different from zero are indicated in the relevant figures using horizontal lines. Cluster-based permutation testing was also applied to performance bias across the angular distance between the presented orientation and the inducer orientation.

To test the significance of our bias analyses, we generated a shuffling distribution after the decoding stage. When trials were sorted based on the relative orientation of the previous trial/stimulus, we randomly sign-flipped this angular distance and re-calculated the bias. We computed the bias for all participants and averaged this score. This process was repeated 10,000 times, and the resulting distribution was z-scored. 

The same z-score transformation was applied to the observed bias score when no random sign-flipping was applied. This z score could then be used to get the (two-tailed) p-value of the original effect relative to the shuffling distribution (in all reported time averages, time points between 250 and 600 ms were used).

All tests were two-sided unless stated otherwise.

Results

Error rates

Participants were accurate in reproducing the target orientation (mean response error 11.73 6 0.70° SEM; mean SD 17.61 6 1.07° SEM; see Table 1 for condition-wise performance). A two-by-two repeated-measures ANOVA on response error showed the main effects of cue type (F(1,19) = 16.49, p, 0.001, h2 = 0.374) and several stimuli presented (F(1,19) = 29.78, p, 0.001, h2 = 0.075). 

Cue type was significant for both one-item or two-item trials, with absolute error higher on report first than on report second trials for both two-item trials (t(19) = 3.972, p, 0.001, d = 0.888; see Table 1) and one-item trials (t(19) = 3.948, Bonferroni-corrected p, 0.001, d = 0.883). By contrast, several items presented primarily affected the report's first conditions. 

Error was higher on report first two than on report first one-item trials (t(19) = 5.665, p, 0.001, d = 1.267) but did not significantly differ between report second two-item and report second one-item trials (t(19) = 1.885, p = 0.075, d = 0.421), leading to a significant interaction between the two factors (F(1,19) = 10.90, p = 0.004, h2 = 0.026). 

Analyses using mixture modeling (Bays et al., 2009) confirmed that errors originating from responses to the noncued grating orientation were rare (swap rate of 0.033 6 0.01 on two-item trials; see also Huang, 2020).

Repulsive performance biases within trials

We analyzed within-trial biases in behavioral performance by assessing whether the reported orientation was systematically reported as closer to or further away from the nontarget orientation on the same trial (see Materials and Methods). 

We restricted the analyses to two-item trials. Figure 2A shows the performance bias for all absolute angular distances between the first and second grating orientation for report first and report second trials. In trials with report first cues, there was no significant bias toward or away from the interfering second grating orientation that was not relevant to the task at hand (t(19) = 0.74, p = 0.467). 

In contrast, trials with report second cues revealed significant biases away from the initially encoded first grating orientation (t(19) = 2.33, p = 0.031; illustrated in Fig. 2B). The repulsive bias in report second trials was confirmed using a cluster-based permutation test, showing a significant cluster (p = 0.012) when the angular distance between the two orientations was between 10° and 49° (Fig. 2A).

Attractive performance bias between trials

We next evaluated the between-trial bias on responses in the current trial toward the orientation that was cued on the previous trial (Fig. 3). We assessed the performance bias as a function of angular differences between the target grating on the current and on the previous trial. 

The analysis also considered the position of the target grating in the current trial (first or second) and the number of items on the current trial (one-item or two-item). For consistency, we term all trials where participants report the first grating orientation report first trials and trials where participants report the second orientation report second trials, regardless of the number of gratings presented. 

Again, we calculated the sum of the bias across angular distances between targets in the current and previous trials (Fig. 3A, B). In contrast to the repulsive bias described in the previous section, we found that all conditions showed an attractive performance bias (all p, 0.05 in two-sided statistical tests). The attractive serial bias was most pronounced for small to intermediate angular distances between the inducer and current orientation (0–60°). 

A repeated-measures ANOVA on the sum of biases across angular distances indicated an effect of cue type, with larger biases occurring in report first trials (F(1,19) = 5.706, p = 0.027, h2 = 0.172), but not of the number of gratings presented in a trial (F(1,19) = 0.980, p = 0.335, h2 = 0.007). The two factors did not interact (F(1,19) = 0.377, p = 0.547, h2 = 0.002). This shows that the bias was stronger when recalling the first item, which was encoded closer in time to the previous trial.

For more information:1950477648nn@gmail.com