Figures and data
Perceptual distractor produces antagonistic biases during recall.
A. Visual working memory task with a distractor. Two randomly oriented gratings were flashed briefly, one in each hemifield. A retro-cue (white semicircle) indicated the location of the erstwhile grating that would likely be probed for response (70% validity). Following a variable delay-period, on 80% of the trials, a singleton perceptual distractor grating was flashed either on the right or left hemifield (upper and lower rows); on the remaining trials (interleaved) no distractor appeared (middle row). At the end of the trial, a central response probe (white or black) indicated the location of the grating (cued or uncued, respectively) to be recalled and reported. B. Distribution (thin histogram) of the orientation report errors for the cued (magenta) and uncued (green) trials (n=24, data pooled across participants). Thick lines: von Mises fits. (Inset) Circular standard deviation (CSD) of the report errors for “no distractor” trials, for cued (x-axis) versus uncued (y-axis) grating reports. Dashed line: line of equality. Data points: individual participants. C. CSD for no distractor trials (black), when the distractor appeared in the same (blue), or opposite (orange), hemifield as the probed memorandum. Violin plots: kernel density estimates; center line: median; box limits: upper and lower quartiles; whiskers: 1.5x interquartile range. D. Change in precision between the distractor-same and distractor-opposite trials (Dk=kDistSame– kDistOpp) for the cued (magenta) and uncued (green) trials. Error bars: Jackknife. E. Schematic of bias. (top) The distractor (left grating) may bias the memorandum’s (right grating) recalled orientation away from its own (left bar), towards its own (right bar), or not at all (middle bar). (bottom) Error in the probed memorandum’s orientation (signed) (y-axis) as a function of its orientation relative to the distractor (x-axis). Positive and negative values (both axes): clockwise and counter-clockwise relative orientations, respectively. Solid and dashed curves: attractive bias (toward), or repulsive bias (away from), the distractor’s orientation, respectively. F. Average behavioral bias (n=24) when the distractor appeared in the same hemifield as the cued memorandum. Solid line: Derivative of Gaussian fit (Methods). Data smoothed over three successive orientation bins for visualization purposes only. Error bars: s.e.m. Gray shaded rectangle: untested relative orientation values (±10°). (Inset, top right) Violin plot: bias strengths across participants; positive (negative) values denote attractive (repulsive) bias. Other conventions are the same as in panel C. G. Same as in panel F, but for trials in which the distractor appeared in the hemifield opposite to the cued memorandum. Other conventions are the same as in panel F. (All panels) *: p<0.05; **: p<0.01; n.s.: not significant.
Stronger mnemonic neural representation mitigates distractor-induced bias
A. Orientation tuning map for memory array decoding for each time point relative to memory array onset. The tuning is computed as the similarity (sign reversed, mean-centered Mahalanobis distance) between the neural activity of the test trial and reference trials for stimuli at different relative orientations (y-axis). (All panels) Unless indicated otherwise, data averaged across participants (n=23, EEG) and for the left and the right gratings. B. (Left) Memoranda decoding accuracy (dot product of tuning curve with cosine, normalized uV, Methods) for each time point locked to the memory array onset (dashed vertical bar). Horizontal black bar: significant decoding epochs (cluster-based permutation test, Methods). Shading: s.e.m. Profiles were smoothed with an 80 ms rectangular moving window for clarity of visualization. (Inset) Posterior electrodes used for the decoding analysis, contralateral to the respective grating. (Right) Same as in the left panel but showing decoding accuracy separately for cued (magenta) and uncued (green) gratings following retro-cue onset (arrowhead). (Inset) Violin plots showing the distribution of average decoding accuracy across participants (n=23) for cued and uncued gratings following the cue onset until the earliest distractor onset (1.2s). C. Behavioral bias curves for the cued memorandum, but plotted based on a median split of memorandum maintenance strength – quantified with neural decoding accuracy during the early delay period (gray shaded bar, panel B, right) – when the distractor appeared on the same hemifield as the memorandum (n=23). Solid curve, filled symbols and filled violin plot (inset): Weaker maintenance trials. Dashed curve, open symbols and open violin plot (inset). Stronger maintenance trials. Other conventions are the same as in Figure 1F. D. Same as in C, but when the distractor appeared on the hemifield opposite to the cued memorandum. Other conventions are the same as in panel C and Figure 1G. (All panels) *: p<0.05; **: p<0.01; ***:p<0.001; n.s.: not significant.
Stronger distractor encoding enhances distractor-induced bias.
A. Same as in Figure 2A, but showing the orientation tuning map for distractor decoding for each time point relative to distractor onset. Data averaged across participants (n=23, EEG). Other conventions are the same as Figure 2A. B. Same as in Figure 2B (right), but showing distractor decoding accuracy for each time point locked to the distractor onset (dashed vertical bar). Magenta and green: Distractor decoding accuracy when it appeared on the cued or uncued hemifields, respectively. Horizontal bars: significant decoding epochs (cluster-based permutation test, Methods). Shading: s.e.m. (Inset, top) Posterior electrodes used for the decoding analysis. (Inset, right) Violin plots of distractor decoding accuracy on the cued (magenta) and uncued (green) hemifields. C. Same as in Figure 2C, but showing behavioral bias curves for the cued memorandum based on a median split of distractor encoding strength – quantified with neural decoding accuracy following distractor presentation (gray shaded bar, panel B) – when the distractor appeared in the same hemifield as the memorandum (n=23). Solid curve and filled violin plot (inset): Stronger distractor encoding trials. Dashed curve and open violin plot (inset). Weaker distractor encoding trials. Other conventions are the same as in Figure 2C. D. Same as in C, but when the distractor appeared in the hemifield opposite to the cued memorandum. Other conventions are the same as in panel C and Figure 2D. (All panels) *: p<0.05; **: p<0.01; ***:p<0.001; n.s.: not significant.
Distractor’s timing and gating predict mnemonic biases.
A. Event-related potentials (ERP) locked to distractor onset (dashed vertical line at 0s) over posterior electrodes contralateral to distractor. Magenta and green traces: ERPs evoked by distractor in the cued and the uncued hemifields, respectively. Solid vertical lines at 0.15s-0.2s and 0.2s-0.35s demarcate the temporal windows for computing N1 and P2/P3a amplitudes, respectively. Shading: s.e.m. (Insets) Distribution of the P2/P3a (top) and N1 (bottom) amplitudes evoked by the distractor in the cued (magenta) and uncued (green) hemifields, respectively. B. Same as in Figure 3C, but showing behavioral bias curves for the cued memorandum separately for trials in which the distractor appeared early (dashed, open symbols) or late (solid, filled symbols) during the delay period, when the distractor appeared on the same hemifield as the memorandum. (Inset) Violin plots show the distribution of bias strengths across participants (n=23) for early (open) and late (filled) distractor trials. C. Same as in panel B, but showing behavioral bias curves for the cued memorandum separately for trials with strong (dashed, open symbols) or weak (solid, filled symbols) distractor-evoked P2/P3a amplitude, a putative marker of distractor gating (panel A). (Inset) Violin plots showing the distribution of bias strengths across participants (n=23) for strong (open) and weak (filled) distractor gating trials. (All panels) *: p<0.05; **: p<0.01; ***: p<0.001; n.s.: not significant.
Two-tier attractor model explains spatially antagonistic, mnemonic biases.
A. Two-tier ring attractor model. The visual cortex (VC, bottom) and higher cortex (HC, top) were modeled with separate ring attractors. Distinct attractors also encoded stimuli in the cued (C, left, magenta) and the uncued (U, right, green) hemifields. Top-down input from HC to VC was stronger for the cued (thick black line) compared to the uncued hemifield. Mutual inhibition between the VC,C and VC,U attractors was mediated by topographic, cross-hemispheric inhibitory connections (black circles with horizontal arcs). (Inset, top) Recurrent excitatory (E-E) connectivity schematic showing local and topographic connectivity profile (see text for details). B. An energy landscape intuition for spatially antagonistic distractor-induced biases. (Top row) The memorandum’s orientation (blue or orange circle) is maintained at the lowest energy state (highest firing rate). (Middle row, left) When the distractor (gray circle) appears in the same hemifield as the memorandum, it causes a dip (energy minimum; solid gray line) in the energy landscape. (Middle row, right) By contrast when the distractor appears in the opposite hemifield, it causes a peak (energy maximum; dashed gray line) in the energy landscape. (Bottom row) As a result, the final energy landscape, obtained by summing the memorandum and distractor energies, attracts the memorandum’s representation towards the distractor’s orientation when the latter appears in the same hemifield (blue circle), whereas it repels the memorandum’s representation away from the distractor’s orientation when the latter appears in the opposite hemifield (orange circle). C. Firing rate heat maps of the neural activity; hotter colors indicate higher activity. y-axis: E neurons arranged in order of orientation preference on the ring; x-axis: Time. (Top) In this example, the onset (dashed vertical line) of the memorandum gratings (0° orientation) causes a localized bump in firing rate at the respective location on the ring, which persists throughout the delay period (gray horizontal bar). Thick black curve: orientation decoded from the population activity vector. When the distractor grating (black triangle; −30° orientation) is presented transiently, and in the same hemifield as the memorandum, the latter’s representation is biased toward the distractor’s orientation. (Bottom) Same as in the top panel, but in this case the distractor grating (−30° orientation) is presented transiently, and in the opposite hemifield as the memorandum (+30° orientation). Here, the memorandum’s representation is biased away from the distractor’s orientation. D. Decoded neural representations across time for six different orientations of the memoranda (curves, distinct simulations). Distractor grating (black triangle; 0° orientation) presented in the same (top, blue curves) or the opposite (bottom, red curves) hemifield as the respective memoranda.
Model simulations qualitatively replicate distractor-induced mnemonic biases.
A. The model’s simulated behavioral bias when the distractor appeared in the same (blue), or opposite (orange), hemifield as the cued memorandum. Compare with Figures 1F-G (experimental data). Other conventions are the same as in Figures 1F-G. B. Same as in panel A, but showing behavioral bias for the cued memorandum plotted separately for simulated trials with weak (solid curve and filled symbols) or strong (dashed curve and open symbols) memorandum maintenance trials (compare with experimental Figure 2C). C. Same as in panel B, but for simulated trials with weak (solid curve and filled symbols) or strong (dashed curve and open symbols) memorandum encoding. D. Same as in panel B, but for simulated trials with weak (dashed curve and open symbols) or strong (solid curve and filled symbols) distractor encoding (compare with experimental Figure 3C). E. Same as in panel B, but for simulated trials with early (dashed curve and open symbols) or late (solid curve and filled symbols) distractor presentation in the delay-period (compare with experimental Figure 4B). F. Same as in panel B, but for simulated trials with weak (solid curve and filled symbols) or strong (dashed curve and open symbols) distractor gating (compare with experimental Figure 4C) (B-F) Other conventions are the same as in Figure 2B. See text for details of how each simulation condition was controlled.
Distractor-induced behavioral biases for the uncued memorandum.
A. Same as in Figure 1G (main text) but showing the average behavioral bias for the uncued memorandum when the distractor appeared in the same hemifield as the former. Other conventions are the same as in Figure 1G. B. Same as in panel A, but for trials in which the distractor appeared in the hemifield opposite to the uncued memorandum. Other conventions are the same as in panel A and Figure 1H (main text).
von Mises mixture model fit and evaluation.
A. Schematic of a mixture model comprising two von Mises and a uniform distribution in which participants responses are modeled as a combination of target (memorandum) and distractor orientation reports, and guesses (“target+distractor”/T+D model). κt and κd denote the precision associated with the target and distractor responses respectively. ⍺ and β denote the contribution of the target and distractor to the orientation responses whereas, γ denotes the contribution of the uniform component (⍺ + β + γ =1). B. Same as in A, but for a mixture model comprising one von Mises and a uniform distribution to model participants’ responses based on target orientation reports (“target alone”/T model) and guesses alone. Other conventions are the same as in panel A. C. Model comparisons. (Left) Difference between corrected Akaike information criterion (AICc) values for the “target+distractor” model fit and “target alone” model fit. A lower (more negative) value indicates a preferred model. Individual points: Participants. Violins, and box and whisker plots follow the same conventions as in Figure 1F-G (insets). (Right) Same as in the left panel but for the Bayesian information criterion (BIC). Asterisks: significant difference of median values between models, based on permutation tests. *p<0.05, **p<0.01, ***p<0.001, n.s.: not significant. D. Scatter showing the distribution of ⍺ (target weight) and β (distractor weight) obtained by fitting participants’ responses with the “target+distractor” combined mixture model. Points: Individual participants. Dashed diagonal line: line of equality.
Effect of memorandum encoding on distractor-induced bias.
A. Same as in Figure 2C (main text), but showing behavioral bias curves for the cued memorandum based on a median split of memorandum encoding strength – quantified with neural decoding accuracy following memorandum presentation (gray shaded bar, Fig. 2B, left) – when the distractor appeared in the same hemifield as the memorandum (n=23). Solid curve and filled violin plot (inset): Weaker memorandum encoding trials. Dashed curve and open violin plot (inset). Stronger memorandum encoding trials. B. Same as in panel A, but when the distractor appeared in the hemifield opposite to the cued memorandum. Other conventions are the same as in Figures 2C-D (main text).
Distractor-opposite and omnibus analysis results.
A. Same as in Figure 4B-C (main text) but showing the behavioral bias curves for the cued memorandum when the distractor appeared on the hemifield opposite to the cued memorandum – for trials in which the distractor appeared early (dashed, open symbols, unfilled violins) or late (solid, filled symbols, filled violins) during the delay period. B. Same as panel A, but for trials with strong (dashed) or weak (solid) distractor-evoked P2/P3a amplitude. (A-B) Other conventions are the same as in Figures 4B-C. C. Same as panel Figure 4B (main text), but showing the results of an omnibus analysis combining across conditions with high distractor interference – late timing, strong encoding, weak gating (solid, filled symbols, filled violins) – or low distractor interference – early timing, weak encoding, strong gating (dashed, open symbols, unfilled violins) – for trials in which the distractor appeared on the same hemifield as the cued memorandum. D. Same as panel C, but when the distractor appeared on hemifield opposite to the cued memorandum. (C-D), other conventions are the same as in Figure 4B (main text) and SI Figure S4A. (All panels) *: p<0.05; **: p<0.01; ***: p<0.001; n.s.: not significant.
Simulations of neural bias and distractor-opposite effects.
A. Simulated neural bias for the cued memorandum when the distractor appeared in the same hemifield (blue) or the opposite hemifield (orange) as the former. Other conventions are the same as in Figure 6A (main text). B-F. Same corresponding panels as in Figure 6B-F (main text), except showing simulated behavioral bias when the distractor appeared in hemifield opposite to the respective memorandum. Other conventions are the same as in corresponding panels in Figures 6B-F.
Error-correcting dynamics in WM.
A. Change in error in the VC attractor readout upon incorporating top-down feedback from the HC attractor. Data shown separately for the cued (magenta) and uncued (green) attractors. Violin plots showing the distribution of values across 10 runs of the model (n=1000 simulations for each run); each simulation initialized with different random seeds, and values were averaged across simulations within a run. B. Change in error obtained by averaging readouts from the VC and the HC attractor relative to the mean readout error for either attractor (average of VC and HC readout errors). Other conventions are the same as in panel A.
