Noise in Competing Representations Determines the Direction of Memory Biases
Curation statements for this article:-
Curated by eLife
eLife Assessment
This valuable study investigates the mechanisms underlying inter-item biases in visual working memory. By experimentally manipulating the relative noise levels of target and non-target items, the authors report bias patterns that are broadly consistent with predictions of their previously proposed normative demixing theory. However, the supporting evidence remains incomplete, as the manuscript lacks a sufficient description of the underlying theory, key assumptions, and a quantitative link between the model and behavioral data. The manuscript would be substantially strengthened by clearer exposition and stronger tests, including analyses of the full error distributions and comparisons with alternative models, which would increase its potential interest to the cognitive neuroscience and computational cognitive science communities.
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
Memories are reconstructions, prone to errors. Historically considered a nuisance, memory errors have recently gained attention when found to be systematically shifted away from or toward non-reported items, promising insights into memory mechanisms. We propose that these biases are optimal and inevitable when the brain disentangles overlapping memory signals, predicting that bias direction depends on the noise distribution between memorized items, not just absolute noise levels. We tested this prediction in four color-memory experiments using novel stimuli with independently varied noise levels. The results support our hypothesis: targets with the same absolute noise level can be repelled from or attracted to non-target items, depending on their relative noise levels. We further show that the model can fit nonlinear bias patterns observed in human data with noise levels as the only free parameters. These findings challenge currently dominant models and support signal disentanglement as a unifying explanation of memory biases.
Article activity feed
-
eLife Assessment
This valuable study investigates the mechanisms underlying inter-item biases in visual working memory. By experimentally manipulating the relative noise levels of target and non-target items, the authors report bias patterns that are broadly consistent with predictions of their previously proposed normative demixing theory. However, the supporting evidence remains incomplete, as the manuscript lacks a sufficient description of the underlying theory, key assumptions, and a quantitative link between the model and behavioral data. The manuscript would be substantially strengthened by clearer exposition and stronger tests, including analyses of the full error distributions and comparisons with alternative models, which would increase its potential interest to the cognitive neuroscience and computational cognitive science …
eLife Assessment
This valuable study investigates the mechanisms underlying inter-item biases in visual working memory. By experimentally manipulating the relative noise levels of target and non-target items, the authors report bias patterns that are broadly consistent with predictions of their previously proposed normative demixing theory. However, the supporting evidence remains incomplete, as the manuscript lacks a sufficient description of the underlying theory, key assumptions, and a quantitative link between the model and behavioral data. The manuscript would be substantially strengthened by clearer exposition and stronger tests, including analyses of the full error distributions and comparisons with alternative models, which would increase its potential interest to the cognitive neuroscience and computational cognitive science communities.
-
Reviewer #1 (Public review):
Summary:
Many previous studies have reported inter-item biases in visual working memory tasks. These biases can be either attractive or repulsive, depending on the particular experiments. It has been difficult to explain these biases in a unifying theoretical framework. Recently, Chetverikov (the first author of the current manuscript) proposed a demixing model for explaining these biases in Ref 22. That paper shows that both attractive and repulsive biases could emerge in the demixing framework depending on the noise properties. The current manuscript seeks to test the predictions of the demixing model experimentally in a series of new experiments and find evidence supporting the demixing model.
Because previous modeling results described in reference 22 (which is a preprint) are essential in interpreting …
Reviewer #1 (Public review):
Summary:
Many previous studies have reported inter-item biases in visual working memory tasks. These biases can be either attractive or repulsive, depending on the particular experiments. It has been difficult to explain these biases in a unifying theoretical framework. Recently, Chetverikov (the first author of the current manuscript) proposed a demixing model for explaining these biases in Ref 22. That paper shows that both attractive and repulsive biases could emerge in the demixing framework depending on the noise properties. The current manuscript seeks to test the predictions of the demixing model experimentally in a series of new experiments and find evidence supporting the demixing model.
Because previous modeling results described in reference 22 (which is a preprint) are essential in interpreting the results reported in the current manuscript, I also studied that preprint and used the results reported in that paper to help interpret the results in this paper. My comments below will also contain discussions of that modeling paper.
Strengths:
Overall, the computational model tested in the paper is novel and interesting.
The demixing framework represents an appealing hypothesis that deserves further investigation.
The current paper provides new empirical data showing that the target stimuli with the same absolute noise level can be either repelled from or attracted to non-target items, depending on the relative noise levels. The observation that biases depend on the relative noise levels is by itself an interesting one, and is consistent with the prediction of the demixing model.
Weaknesses:
While this manuscript contains interesting new experimental observations and theoretical ideas, it has several substantial problems in its current form, which limit the conclusions that can be drawn. The description of the computational model is too brief. The key modeling assumptions need to be better motivated and explained. As the computational models generate different predictions in different regimes, it is a bit difficult to evaluate how well the experimental data support the model at a more quantitative level. Also, the results focused on studying the biases in the behavior; it is unclear whether the model can fully explain the behavior data (such as error distributions or behavioral precision).
Major concerns:
(1) Concerns/suggestions regarding the computational modeling
The current paper seeks to test the predictions of the demixing-based computational model proposed in reference 22. There are several problems with the modeling component in the current paper.
(1a) The description of the model is too brief and difficult to understand. Although the model was proposed in reference 22, it would still be beneficial to provide more details of the model so that readers can understand and appreciate the strengths/limitations of the model.
The generative model and the inference procedure could be better explained to better link the model to the behavior. In particular, how was the observer's behavioral report in each trial modeled? This requires more explanation because currently the demixing procedure estimates four parameters for a given trial, yet for a given trial, only one behavioral report was produced (e.g., current Experiment 1), or two reports were produced sequentially (e.g., current Experiment 2).
(1b) Key modeling assumptions need better justification.
One such key assumption is that on a given trial, each stimulus triggers many samples (or approximately, an entire response distribution), rather than a single sample. This assumption deviates substantially from prior work on ideal observer models. It was not clear whether this assumption is realistic. For the type of stimuli used in the current experiments, perhaps one can argue that each pixel corresponds to one sample of brain activity, thus collectively each stimulus should trigger many samples of activity in the brain. If this were to be the case, it would have two implications. First, the noise parameter in the model should be directly related to the magnitude of the stimulus noise. Thus, one should be able to plug these experimentally-controlled parameter values into the model to directly generate predictions about the biases. Second, when using stimuli with no stimulus variability (e.g., simple grating stimuli), the predicted biases should change. However, it wasn't clear whether this would hold experimentally, i.e., using gratings would lead to different biases or no biases.
If the variability of the samples for a given stimulus involves neural noise, it would be useful to justify why it is reasonable to consider that many samples were generated per stimulus.
(1c) As mentioned in (1b), the model assumes that on each trial, a large number of samples was generated. It would be useful to study and report how the prediction would change when the number of samples generated per stimulus is small. In particular, what happens when each stimulus only generates one measurement? This might be useful for interpreting previous experiment results with grating stimuli.
(1d) Reference 22 studies how the predicted biases depend on the d-prime of the identifying dimension and found that the pattern of the biases varies substantially depending on the information available for the identifying dimension. However, the current paper didn't really discuss this important point. It is also unclear what parameters the authors used for the d-prime of the identifying dimension. Was it fitted directly to the data? The Methods section has some description on the "identifiability dimension", but it was a bit obscure.
Intuitively, when the d-prime of the identifying dimension is very large, the demixing problem becomes irrelevant. In this case, there should not be any biases induced by demixing. In the case of the d-prime for the identifying dimension is 0, the problem should reduce to the simplified 1-d problem studied in reference 22. If my reading of reference 22 was correct, they reported different conclusions. It would be useful to clarify these points.
In any case, the d-prime of the identifying dimension appears to be a key parameter. It would be great to constrain this parameter using the empirical data. When the d-prime of the identifying parameter is small, the observer would easily confuse the probed stimulus with the other stimulus in a given trial. This should lead to poor task performance. Thus, it may be possible to directly estimate the value of the d-prime of the identifying dimension based on the observer's performance, and then use this parameter to generate model predictions accordingly.
(1e) The current model assumes that a large number of samples are generated per stimulus and the brain can manipulate this information to perform the demixing task. It was well documented that visual working memory has a capacity limit (i.e., it can only hold information about a few items); this discrepancy needs to be clarified or addressed.
(2) How well the computational model can explain the experimental data remains not entirely clear
The authors show that there exists a parameter regime that can qualitatively explain the experimental finding. They also show that it is possible to fit the model to the data to explain the bias patterns. However, given that the model is flexible, it would be stronger if the authors could show that the same parameters that explain the biases could also explain other aspects of the behavior, for example, the magnitude of the errors.
In other words, the model is not well constrained in the way it was tested in the paper. But it should be possible to improve it. First, if the noise parameter in the model is determined by the stimulus variability, one can determine it directly based on the external noise in the stimuli (discussed also in 1b) and see what prediction it leads to. Second, from the behavioral data, it may be possible to estimate the noise for the identifying dimension. Doing so will help better constrain the model.
It would also help if the authors could report the best-fitted parameters from the experimental data. From these parameters, one can simulate synthetic data and apply the demixing model to see if the error distribution of the simulated observers is indeed similar to the experimentally measured error distribution. That way, one can check whether the fitted parameter explains the observer's behavioral performance beyond the biases.
Other comments:
(1) How does the model account for the swap errors? I am not sure I understood the way how the swap errors were treated in the paper. To me, substantial swap errors seem to be a consequence of having low d-prime values for the identifying dimension; that is, if there is only little information to discriminate the identity of the two stimuli, swap errors would be large. However, this possibility didn't seem to be mentioned in the paper.
(2) Since the solution of the demixing problem was obtained using a numerical procedure based on EM. It would be useful to check whether the initialization has affected the biases obtained.
-
Reviewer #2 (Public review):
Summary:
This manuscript investigates the origins of inter-item biases in visual working memory. The authors proposed a computational model where overlapping memory signals are disentangled, inducing memory biases that depend on relative noise levels across items. The key theoretical advance is the prediction that bias direction depends not only on absolute memory noise but on the relative noise levels of target and non-target representations. Using four experiments with color mosaics whose color variability manipulates memory precision, the authors report that biases reverse as a function of relative noise in a manner predicted by the model.
Strengths:
The manuscript is clearly written and theoretically motivated. The experiments are well designed and provide converging evidence for a distinctive and …
Reviewer #2 (Public review):
Summary:
This manuscript investigates the origins of inter-item biases in visual working memory. The authors proposed a computational model where overlapping memory signals are disentangled, inducing memory biases that depend on relative noise levels across items. The key theoretical advance is the prediction that bias direction depends not only on absolute memory noise but on the relative noise levels of target and non-target representations. Using four experiments with color mosaics whose color variability manipulates memory precision, the authors report that biases reverse as a function of relative noise in a manner predicted by the model.
Strengths:
The manuscript is clearly written and theoretically motivated. The experiments are well designed and provide converging evidence for a distinctive and non-intuitive prediction of the proposed model. I found the central result compelling: independently manipulating target and non-target noise leads to qualitatively different bias patterns, consistent with the model's prediction that relative noise is a key determinant of bias direction.
Weaknesses:
The main limitation is that the evidence establishes consistency of the data with the proposed Demixing Model, but does not demonstrate that the model provides a unique explanation of the data. Although the manuscript argues that dominant theories struggle to account for the observed reversals, no formal comparison with alternative computational frameworks is presented. In addition, model fitting results are reported only briefly, making it difficult to evaluate fit quality at the level of individual observers.
-
