Changes to information in working memory depend on distinct removal operations

Holding information in working memory is essential for cognition, but removing unwanted thoughts is equally important. Here we use multivariate pattern analyses of brain activity to demonstrate the successful manipulation and removal of information from working memory using different strategies including suppressing a specific thought, replacing a thought with a different one, and clearing the mind of all thought. These strategies are supported by distinct brain regions and have differential consequences for allowing new information to be encoded.

. Summary statistics of the WM operation classifiers. The classifier was trained with and tested on the central study data (k-fold leave-one-out cross-validation). For each operation, the statistics for the classifier accuracy (row 1) and AUC (row 2) are summarized in the mean ± SEM, t-value (DOF = 49) of one-sample T-tests (two-sided) with chance level (1/the number of operations for accuracy, 0.5 for AUC), p-value, effect size (Cohen's d), and 95% confidence interval in order.
Between-subject classification (4-operation, anatomically aligned) For each class (category or subcategory), the statistics for the cross-validation (row 1: classifier accuracy, row 2: AUC) and central study decoding (row 3: classifier accuracy, italicized) are summarized in the mean ± SEM, t-value (DOF = 49) of one-sample T-tests (two-sided) with chance level (1/the number of operations for accuracy, 0.5 for AUC), p-value, effect size (Cohen's d), and 95% confidence interval in order. The crossvalidation classifiers were validated with both individualized penalty (see Methods) and a single global penalty.

4-operation classifier
The individualized penalty was applied on the classifiers that tested the central study data.
Supplementary Table 3. Summary statistics of the WM representation trajectories in the category-level (classifier) and item-level (RSA, Fig. 4). The data were evaluated for five 1.38 s (3 TR) time windows beginning at the onset of the manipulation instruction at 2.76 s after stimulus onset and extending to 9.66 s post-stimulus onset (window 1-5). The one-sample T-test (vs. baseline = 0, two-sided, FDR corrected) results for the adjusted decoding value (removal -maintain; category-level | item-level) of each operation and the paired T-test (two-sided, FDR corrected) results across the operations are summarized in t-value, p-value, effect size (Cohen's d), and 95% confidence interval are summarized in order (Fig. 4B). Only the results from the significant windows are reported (-: n.s.). The one-sample T-test (vs. baseline = 0, two-sided) results for the encoding fidelity (same-minus-different) of each operation and the paired T-test (two-sided, Tukey-Kramer corrected) results for the three different encoding fidelities (same-minus-different, same-category, different-category) across the operations are summarized in t-value (DOF), p-value, effect size (Cohen's d), and 95% confidence interval in order (* denotes P < .05, Fig. 5).

Supplementary
Supplementary Table 5. The first contrast examined activation when the item needs to be manipulated (replace, suppress, clear) as compared to when no such manipulation was required (maintain). As in our prior study 2 , this contrast mainly revealed activation in posterior regions of the brain, with relatively sparse activation in prefrontal cortex associated with cognitive control. Supplementary Table 6. The second contrast examined activation when an item had to be removed from working memory (suppress, clear) as compared to when an item remained in working memory (maintain, replace). Once again, the results were consistent with our prior report 2 . There was activation in several different regions involved in cognitive control, but unlike the prior contrast, these included many prefrontal regions, including anterior DLPFC. Supplementary Table 8. The contrast of suppress > clear revealed that the suppress condition revealed much more activation across a broad swath of lateral prefrontal cortex, consistent with our conclusion that the suppress condition is characterized by an active removal process (i.e., one that involves cognitive control). Alternatively, the contrast of clear > suppress revealed activation in the precuneus, as well as regions in the right hemisphere occipital regions adjacent to those that according to Neurosynth (https://neurosynth.org/) are significantly association with the term distraction. This finding is also consistent with our multivariate results suggesting that the clear condition works by removing information from the focus of attention in working memory, but that it does not necessarily influence the memory representation itself.  Table 9. We also compared the suppress and maintain condition, which showed significantly different neural signature for the cognitive process but similar engagement in terms of holding representation in WM. The contrast of suppress > maintain showed greater activation in inferior frontal and opercular regions, which were involved in cognitive control. On the other hand, the contrast of maintain > suppress revealed activation in regions of the ventral visual stream suggesting that maintain requires holding specifics item details more than suppress.