Neural bases of space-specific distractor biases in visual working memory
Curation statements for this article:-
Curated by eLife
eLife Assessment
This useful study combines behavioral reports, EEG decoding, and computational modeling to address an interesting question: how delay-period distractors bias working-memory representations, and how these effects depend on target relevance, distractor location, and the strength of memory maintenance and distractor encoding. However, the supporting evidence is incomplete, as several key claims require clearer statistical tests, better integration of the behavioral and neural results, and more careful consideration of alternative explanations. Stronger engagement with prior literature would also substantially strengthen the manuscript and increase its potential interest to researchers in systems, cognitive, and computational neuroscience.
This article has been Reviewed by the following groups
Discuss this preprint
Start a discussion What are Sciety discussions?Listed in
- Evaluated articles (eLife)
Abstract
Information held in working memory (WM) is remarkably resilient to distraction. Yet, perceptual distractors that share mnemonic features can impact WM profoundly; the neural basis of this phenomenon remains unclear. With multivariate decoding of human electroencephalography recordings, we investigate how delay-period perceptual distractors bias WM. Participants memorized the orientations of cued and uncued grating memoranda that appeared in opposite hemifields. A grating distractor, flashed during the delay period, produces space-specific biases: memorized features are attracted towards or repelled away from the distractor’s orientation depending, respectively, on when the distractor appeared in the same hemifield as the memorandum, or opposite to it. Neural prioritization in WM by cueing, and stronger memorandum maintenance mitigate this bias, whereas stronger distractor encoding enhances it. Lastly, a ring-attractor model with cross-hemifield inhibition mechanistically explains the origins of these spatially-antagonistic biases. Our results reveal how lateralized sensory buffers critically enable perceptual distractors to bias visual WM.
Lay Summary
Working memory (WM) – the ability to momentarily store important items – is remarkably robust to distraction. Yet, when a salient object with features resembling the memorized item (“perceptual distractor”) appears in the environment, it can alter WM appreciably. What neural mechanisms render WM resilient to distraction, and how can perceptual distractors affect it so profoundly? We address this question by presenting a salient distractor, at unpredictable times, during the WM delay-period. Surprisingly, the distractor’s bias on WM depends on its proximity to the memorandum’s original location. With state-of-the-art neural decoding and computational modeling, we identify mechanisms that mediate these space-specific distractor effects. The findings advance our understanding of WM’s resilience to distraction and may inform cognitive therapies for treating WM deficits.
Article activity feed
-
eLife Assessment
This useful study combines behavioral reports, EEG decoding, and computational modeling to address an interesting question: how delay-period distractors bias working-memory representations, and how these effects depend on target relevance, distractor location, and the strength of memory maintenance and distractor encoding. However, the supporting evidence is incomplete, as several key claims require clearer statistical tests, better integration of the behavioral and neural results, and more careful consideration of alternative explanations. Stronger engagement with prior literature would also substantially strengthen the manuscript and increase its potential interest to researchers in systems, cognitive, and computational neuroscience.
-
Reviewer #1 (Public review):
Summary:
In this study, Deepak V. Raya and colleagues combined behavioral measures with EEG recordings to investigate how distractors presented during the working memory delay influence memory representations. Using oriented gratings as stimuli and a continuous estimation task, the authors systematically manipulated factors that may modulate distractor interference, including the behavioral relevance of the WM item (cued vs. uncued) and the spatial relationship between the distractor and the WM item. By analyzing the relative orientation between the WM item and the distractor, the authors showed that distractors presented at the same location as the WM item induced an attractive bias (i.e., reported orientations biased toward that of the distractor), whereas distractors presented at the opposite location …
Reviewer #1 (Public review):
Summary:
In this study, Deepak V. Raya and colleagues combined behavioral measures with EEG recordings to investigate how distractors presented during the working memory delay influence memory representations. Using oriented gratings as stimuli and a continuous estimation task, the authors systematically manipulated factors that may modulate distractor interference, including the behavioral relevance of the WM item (cued vs. uncued) and the spatial relationship between the distractor and the WM item. By analyzing the relative orientation between the WM item and the distractor, the authors showed that distractors presented at the same location as the WM item induced an attractive bias (i.e., reported orientations biased toward that of the distractor), whereas distractors presented at the opposite location produced a weaker effect, with any systematic bias tending to be repulsive. Through a combination of behavioral analyses and EEG-based decoding, the authors further examined and revealed factors that modulate the magnitude of distractor interference, including cueing status, the strength of memory maintenance, distractor timing, and neural indices of distractor encoding and gating. Lastly, the authors propose a computational account of these effects by implementing a two-layer ring attractor model that captures several key behavioral patterns observed in the data.
Strengths:
The influence of distractors on working memory has been extensively studied both behaviorally and with neuroimaging. The present study advances this literature by providing a more comprehensive account that jointly manipulates and quantifies many key factors, including cueing (behavioral relevance), the spatial relationship between WM items and distractors, and distractor timing. This integrative approach enables a more systematic characterization of how different sources of interference interact. A particular strength of the study is the use of EEG combined with multivariate decoding to track the dynamics of memory and distractor representations. Compared to prior fMRI work, this approach provides a time-resolved view of how encoding, maintenance, and distractor processing unfold over time. This is especially valuable for dissociating memory maintenance and stimulus encoding, or gating contribute to behavioral interference, which is more difficult to achieve with fMRI.
Behaviorally, while most previous studies have reported attractive biases by distractors, the current study identified a repulsive effect when distractors were in the opposite hemifield from the WM item. Overall, the study provides a rich investigation of distractor interference in working memory and will be of interest to researchers studying the neural and computational mechanisms that protect memory representations from distraction.
Weaknesses:
(1) In the paragraph starting around line 125, the authors reported a 2-way ANOVA (cue/uncued × same/opposite side) restricted to trials in which a distractor was present. However, the subsequent post-hoc analyses compared distractor-present trials (same or opposite side) with no-distractor trials, which were not included in the ANOVA. While both analyses were informative, presenting them together in this way was somewhat confusing, as the post-hoc tests extended beyond the factors and conditions analyzed by the ANOVA. I suggest presenting these analyses separately and clarifying their distinct purposes. Additionally, Figure 1C appeared to reflect only the pairwise comparisons; including a figure that directly visualizes the two-way ANOVA results would improve clarity.
(2) In lines 138-150, the authors fitted von Mises functions to the distributions of memory error and reported that the effect of distractor location (same vs. opposite) was stronger in the uncued condition than in the cued condition. However, this result appears difficult to reconcile with the earlier 2-way ANOVA, which showed no interaction between cueing and distractor location. It is unclear whether this discrepancy arose from differences in the dependent measures (CSD vs. κ), statistical procedures, or other factors. Clarifying how these two sets of results should be interpreted together would improve the clarity of the findings.
(3) For the analyses in Figures 1B and 1D, parametric functions were fitted to the distributions of memory error using aggregated data. Models of memory error distributions have been central to ongoing debates in the working memory literature (e.g., Schurgin, Wixted, & Brady, 2020; van den Berg, Awh, & Ma, 2014). Fitting functions/curves to aggregated data can be problematic, as it distorts the underlying distributions at the individual level. I suggest performing the fits on the individual data and analyzing the fitted parameters across participants using appropriate group-level statistical tests.
(4) At the end of the first Results section (lines 234-235), the authors concluded that cued memoranda were "better shielded from interference" than uncued memoranda. However, I did not see a clear statistical test directly supporting this. This statement appeared to rely mainly on Figure 1D, which showed a stronger location effect (same vs. opposite) when the memory item was uncued. However, this analysis does not directly test whether distractors impair uncued items more than cued items overall. Supporting this broader claim would require a direct comparison of distractor effects (e.g., distractor vs. no-distractor) between cued and uncued conditions, or an interaction test involving cueing and distractor presence (e.g., either by pooling different distractor locations, or focusing on the same-location condition if opposite-location distractors show no significant effect).
(5) While the attractive and repulsive biases are an interesting finding, it was demonstrated only at the behavioral level. It would be informative to examine whether the biases are reflected in the decoding results. For example, after deriving trial-wise orientation tuning functions, one could estimate decoded orientations (e.g., via vector averaging or the peak of the tuning curve) and assess bias at the neural level. Although EEG SNR may limit recovery of full function of the memory error (e.g., Figure 1F-G), grouping trials into fewer bins (even with just two bins) may still allow detection of the overall direction of the bias in the decoding results. This type of decoding bias has been reported in other contexts (GY Bae - NeuroImage, 2021).
(6) The analysis P2/P3a requires more explanations. Typically, these components are extracted from trial-averaged ERP. The methods section also mentioned "averaged across channels and trials to obtain the ERP waveform." However, to split the trials, these components have to be identified at a single-trial level. More details are needed in the Methods.
(7) Components such as P3a are often linked to attentional capture and orienting, which would predict increased, rather than decreased, distractor interference. The interpretation of this signal as reflecting gating appears to be inferred from the observed relationship between larger P3a amplitudes and weaker interference. The N2pc component is a well-established index of spatial attention allocation and may be particularly relevant (and useful) here, given the lateralized distractor. Have the authors tested whether distractor-evoked N2pc can be used to split trials and examine its relationship with the bias?
(8) Line 676 in the Discussion states "possibly by error-correcting top-down control mechanisms." It is unclear which results provide support for this interpretation, except that there are stronger feedback connections at the cued location in the attractor ring model.
-
Reviewer #2 (Public review):
Summary:
Understanding the factors and mechanisms underlying the deleterious effects of distraction, and protection from distraction, in working memory is an important question that has a long and rich history in psychology and neuroscience, and continues to be highly relevant. In this study, the authors recorded the EEG while subjects viewed the initial presentation of two oriented-grating stimuli, aligned on either side of fixation along the horizontal meridian (memory array), followed by a 70%-valid cue, then one of three distractor conditions (overlapping cued item (40%), opposite cued item (40%), no distraction (20%)), followed by recall ("delayed estimation"). The behavioral and EEG results from this procedure are complemented with computational modeling with a two-tier bump-attractor model.
Weaknesses:
Reviewer #2 (Public review):
Summary:
Understanding the factors and mechanisms underlying the deleterious effects of distraction, and protection from distraction, in working memory is an important question that has a long and rich history in psychology and neuroscience, and continues to be highly relevant. In this study, the authors recorded the EEG while subjects viewed the initial presentation of two oriented-grating stimuli, aligned on either side of fixation along the horizontal meridian (memory array), followed by a 70%-valid cue, then one of three distractor conditions (overlapping cued item (40%), opposite cued item (40%), no distraction (20%)), followed by recall ("delayed estimation"). The behavioral and EEG results from this procedure are complemented with computational modeling with a two-tier bump-attractor model.
Weaknesses:
Interpretation of the results is complicated by several factors. One is the non-consideration of a considerable amount of extant research that is highly relevant to the question of interest (these include seminal studies from Gi-Yeul Bae and from Tatiana Pasternak). Relatedly, the manuscript emphasizes biasing effects of distractors to the exclusion of a conceptually distinct effect: degradation of representational precision. (For example, the actual focus of the study of Wimmer et al. (2014) that the manuscript cites with reference to bias is the degradation of precision; one only has to read the title of this paper to know this.) Also relatedly, the authors are aware of the possibility of misbinding (a.k.a. "swap") errors, in which subjects mistakenly recall a high-fidelity representation of a foil (in this case, the distractor) rather than the target, but they (1) fail to cite any of the extensive literature on this topic and (2) seem to erroneously attribute what their analyses would seem to identify as misbinding errors as "antagonistic bias" exerted by the distractor on the target item.
A second concern relates to the interpretation of patterns in the empirical results. In particular, Figure 1G is interpreted as displaying a pattern of repulsive bias exerted by the distractor on trials when the distractor appeared at the location opposite to the cued item. However, it is not clear that this is a repulsive bias. Rather, what the plot shows is that report error is attracted to "near" distractors with a positively signed offset but repelled by "near" distractors with a negatively signed offset. Stated another way, when one applies a model-free assessment of the influence of the distractor on the memorandum, there is no systematic bias: the AOC of positively signed offset values from 0 to +45 deg is roughly the same as the AOC negatively signed offset values from 0 to -45 deg. The same also seems to be true, albeit with a smaller magnitude, for trials featuring "stronger mnemonic neural representation" that are illustrated in Figure 2. And so it's unclear that the effect of the "Dist. Opp" distractor is indeed a repulsive bias, rather than a loss of precision.
The third primary concern is that the results from simulations from the two-tier bump-attractor modeling are difficult to interpret due to several poorly motivated and seemingly "hand-coded" assumptions. These include the (seemingly arbitrary) strengthening of HCVC feedback connections by 20% for cued vs. uncued items; and the choice to "transiently block[ed] feedforward connections from the VC to the HC during the maintenance epoch" as a consequence of cuing. There is frankly no evidence that the latter phenomenon actually happens in primate brains performing comparable tasks, including in papers (such as from Xu and from Rademaker) that are cited in this manuscript. The current consensus is that priority-related rotations of representational geometry are the scheme employed by mammalian nervous systems to control the otherwise deleterious effects of distraction.
-
