We appreciate the reviewer's positive assessment and their suggestions. We address their specific questions and weaknesses below:

• Small dataset and limited N-back window We appreciate the reviewer highlighting the limited size of our stimuli and task sets. We would like to clarify that our stimuli dataset includes 4 categories, each with 2 identities. For each identity, there are approximately 20 view angles in the training set and 4 view angles in the validation set. Additionally, we include 2 additional identities for each category in the validation novel object dataset, resulting in a total of 192 stimuli. Despite this dataset size, the model is capable of generalizing to identities with novel angles and novel identities, suggesting that the model is indeed learning (and generalizing) the task. We argue that with larger datasets, the model should retain the same representation geometry as discovered with the given dataset.
  To directly address the reviewer's concerns, we ran several additional experiments. In particular, we trained models on a relatively larger dataset containing 8 categories and 4 identities per category. Due to rendering quality restrictions, we obtained models with comparable training and validation accuracy (~80%), suggesting that the models still captured the task dynamics. We performed a cross-condition decoding analysis (similar to Fig. 2a) and found consistent results, indicating shared representation for vanilla RNNs but not gated RNNs. The relatively low decoding accuracy is a direct consequence of the lower saturation accuracy of the model that will likely be ameliorated with additional training time. Additionally, we further trained models to perform 1-4 back tasks across all 3 features (adding the additional 4-back task setting). Models achieved comparable performance to our original 9-task model, and we were able to replicate the findings from Fig. 4f and g (see attached 1-page PDF Fig 2a). Together, these new results suggested that our findings generalize to broader settings including more diverse object categories and broader task settings.
• Suggestion to add realistic background We thank the reviewer for their suggestion and acknowledge that more natural experimental settings are desired and may be important for proper analysis of the neural computations underlying working memory. To address the reviewer’s concern, we ran an additional experiment using stimuli overlaid on synthetically generated natural textures. For this, we synthesized textures as stimulus background using the method from Texture Synthesis and overlaid the 3D objects on them to create each frame (see attached 1-page PDF Fig 1c). We then trained models to perform n-back tasks with these new stimuli. We found that models achieved equally well performance levels with comparable speed (see attached 1-page PDF Fig 1d). In other words, we found that the choice of background does not affect our main results. We plan to include a more detailed analysis of this model in the revised submission (with the full set of analyses).
• Interpretation of subsection 4.3 Please refer to the general reply for an updated interpretation.
• Does the training transfer from Imagenet to Shapenet? As mentioned on line 122-123, our vision backbone consisted of an Imagenet-trained ResNet50 model with fixed parameters. In all our experiments, we used the unit activations of this model as the input to the RNN models. All object features including category, identity, and location were highly decodable from these activations (category: 100.00%, identity: 99.57%, location: 100.00%; 2-fold cross-validation), confirming that the Imagenet trained ResNet50 model generalized to our stimulus set from ShapeNet.
• Orthogonalization index interpretation We agree with the reviewer that the high dimensionality of the latent space, specially the perceptual space, could make the comparison potentially invalid. To address the reviewer’s concern, we repeated this analysis by first applying PCA on the activations within perceptual and RNN spaces and equalized the number of dimensions in both. We then calculated the orthogonality measure on the dimensionality matched spaces (perceptual and RNN encoding). This analysis replicated the same findings from the original analysis done without dimensionality matching (see Fig. 1e in the attached 1-page PDF). Lastly, we want to emphasize that the point made in our orthogonality analysis experiments is the difference in orthogonality between the two spaces and we do not directly examine the absolute value of orthogonality, which as the reviewer suggested may be affected by the dimensionality of the latent space.
• Rationale behind using the Procrustes analysis We were interested in investigating whether the geometry of object representations is unchanged across different time steps (e.g. between encoding representation and 1st, 2nd, or 3rd memory representations). For this we needed to analyze the likeness of the representational geometry across time steps and for encoding and memory representations. The orthogonal Procrustes analysis is a statistical shape analysis which enables discovering simple rotation transformations that superimpose a set of vectors/points onto another set. We used this analysis to inquire whether each set of object feature decoders can be rotated to align with the set of same decoders at a different time or for a different stimulus.

We thank the reviewer for recognizing the importance of our study. We have carefully considered the reviewer's comments and have provided the following response:

• Typos. We sincerely thank the reviewer for carefully listing all of the typos in our original draft. We will fix them and more thoroughly review the revised manuscript to ensure clarity.

• Missing details about model training. Thank you for highlighting this. We have included additional methodological (e.g., training details) in our general response to all reviewers. In the revised version of our manuscript, we will include those details, in addition to other missing method details.

• Discussion on the relevance of generalization to novel object/view angle. The primary reason for our evaluation of novel view/angle generalization was to highlight that the mechanisms of working memory learned by the model were not strictly limited to the specific set of visual features used during training. Hence, we used the same stimuli but with novel views/angles. Nevertheless, we agree with the reviewer that it will be potentially interesting to further examine the representational dynamics of RNNs when considering alternative stimuli (e.g. novel objects). (Note, however, that novel objects would prevent us from evaluating on the n-back identity task, since this requires the same stimulus.) Given the limited time at our disposal for providing a response to the reviews we were unable to perform further analyses in this direction. However, we have added this suggestion as an important future direction in our discussion section.

• Interpretation of the orthogonalization results. Regarding the update of the orthogonalization analysis and interpretation, we kindly ask the reviewer to refer to our general reply for a detailed explanation due to the limited character count for individual rebuttals.

• Causal relevance test. We thank the reviewer for their suggestion. We performed a new analysis to address this point. Specifically, we considered trials from the 1-back location task where two objects had matching location property. We then manually adjusted the hidden state of the network in the direction normal to the decoder’s decision boundary. We then calculated the network’s generated output probabilities as a result of this hidden state adjustment. This analysis revealed that when adjusting the hidden state such that the decoder judges the object location to be different from its original value, the probability of match output is reduced while the probability of non-match and no-action outputs increase (Fig 1c in the attached 1-page PDF). This result thus supports the causal role of the particular encoding of object properties on the network’s generated behavioral output.

• Within time-step decoding analysis. Following the reviewer’s suggestion, we performed the decoding analyses for each time step separately. The results qualitatively mirrored those we had reported in our Procrustes analysis that showed near perfect decodability up the time step where the object information was necessary for performing the task. Please see Fig. 1b in the attached 1-page PDF. In addition, as suggested by the reviewer, we acknowledge that given the near-perfect decoding accuracies, using a continuous measure like distance from the decoding hyperplane could provide more nuanced insights into the representation dynamics. We will include the results of this analysis in the final revised version of the manuscript.

• Interpretation and relationship to previous literatures. Our goal from the present work was to characterize the way multidimensional object properties are represented in the RNN models of WM and to reveal the mechanisms employed by these models for concurrent encoding, retention, and recall of information according to task demands. We will revise the abstract and the main paper to more clearly reflect these aims and to highlight its importance and distinctness from prior work. Furthermore, our findings in section 4.3 provide evidence against one of the prominent models of working memory, the slot-based model of working memory (Luck and Vogel 1997, Whittington et al. 2023). Relatedly, our results further show that within the family of n-back tasks, the memory representation is not “sustained” as is commonly suggested by prior studies. We expect that a sustained or dynamic memory representation will largely depend on the task structure (e.g., n-back task vs. a working memory delay with a static fixation).

Luck, S. J., & Vogel, E. K. (1997). The capacity of visual working memory for features and conjunctions. Nature, 390(6657), 279-281.

Whittington, J. C., Dorrell, W., Behrens, T. E., Ganguli, S., & El-Gaby, M. (2023). On prefrontal working memory and hippocampal episodic memory: Unifying memories stored in weights and activity slots. bioRxiv, 2023-11.

• Limited discussion. We agree that the discussion on the limitations of our work was largely incomplete. To address the reviewer’s concern, we substantially revised the Discussion and Conclusion sections, and included a list of limitations and future directions. This includes additional experiments, including the reviewer’s suggestion on studying other working memory tasks. In particular, we are planning to perform an additional analysis using a simple delayed match-to-sample task to compare how working memory representations are maintained in the absence of incoming stimuli (i.e., the N-back task has incoming stimuli at every ‘delay’ period). If indeed the rotation-like dynamics are employed to protect against interference in the N-back task, we hypothesize that there may be the lack of rotation-like dynamics across time in a delayed match-to-sample (in which the delay is just a fixation). Though we could not perform this analysis during this rebuttal period, we are aiming to include this analysis in the final version of this paper.

I thank the authors for their rebuttal, and apologize for the delay in responding. Generally, their responses to my points were helpful in clarifying my doubts. I only have further comments on the following points:

• Orthogonalization results. As I had written in my initial review, I wasn't sure how to interpret the orthogonalization finding in the first place, so I also don't know how to interpret the updated finding of lower orthogonalization in the RNN. It is indeed counterintuitive, as I would think that training with a task would lead to more orthogonalization along task-relevant dimensions. Either way, I don't think this finding is central to the paper's message, so I believe you could simply remove it.

• Causal relevance test. I do appreciate that the authors have run this analysis. However, given the high level of variability in the network's response across trials, I do not think any conclusion can be made from this analysis about the causal relevance of the network's subspaces. The significance of the trends shown in Fig. 1c of the attachment could be simply checked using a measure of correlation between percentages of shift and probability of each given response. I doubt that any of the correlations shown in Fig. 1c will be significant. One possible reason for the high variability might be that the authors subsampled "match" trials and tried to push them towards giving a "mismatch" response. This means that the network's state was pushed towards the "all" direction of the "one vs. all" classifier, which probably does not correspond to any particular category, hence the high variability. It is possible that trying to push a "mismatch" response towards a "match" response (i.e. pushing the network towards a specific category) would lead to more reliable results. However, my intuition might be wrong and I understand that the authors don't have the time to run this analysis. If the authors wish to include this additional analysis as it is, then, they should describe it as inconclusive rather than try to infer any causal effect from it.

• Within-timestep decoding. The finding of constant decoding accuracy for task-relevant features (and slightly degrading accuracy for task-irrelevant ones) is quite straightforward, and I think it would make a nice addition to the paper. In particular, it provides a nice contrast to findings in RNNs that had to categorize objects without any memory requirements, such as [1, 2], in which decoding accuracy was found to be increasing across time. This suggests a potential distinction between dynamic encoding strategies in perception and working memory. One thing that is not fully clear to me: does Fig. 1b in the attachment indicate that the accuracy was still close to 100% also several timesteps after the time of the response? If so, this would be interesting to highlight. To make this figure clearer, I would suggest adding vertical lines corresponding to the response times in the different conditions, so that it is clear when the accuracy is being measured before or after the response.

• Relation to other models. I thank the authors for their clarification. While I'm familiar with the general idea of slot-based models, I haven't had the time to go through the more explicit model proposed by Whittington et al. (2023). However, from a quick skim of that paper it seems like they do address temporally-structured tasks as well (beyond just delayed response)? Particularly, in their Figure 5, they describe a model that shows some similarity to what the authors propose here, whereby stimuli are encoded in a given slot and then rotated, such that the initial slot can be occupied by the new stimulus to be encoded. I would appreciate it if the authors could clarify the distinction between their model and the one in Whittington et al. (2023). In general, I agree about the importance of across-tasks comparisons, and I am glad that the authors plan to add a comparison with a delayed match-to-sample task. I think that will be a nice addition to the paper.

To conclude, as the authors have clarified several points but not made major changes to the paper, I plan to keep my score. Again, apologies for not leaving enough time for more discussion with the authors, but I believe any further revisions would have been minor anyway.

[1] Spoerer, C. J., Kietzmann, T. C., Mehrer, J., Charest, I., & Kriegeskorte, N. (2020). Recurrent neural networks can explain flexible trading of speed and accuracy in biological vision. PLoS computational biology, 16(10), e1008215.

[2] Thorat, S., Aldegheri, G., & Kietzmann, T. C. (2021). Category-orthogonal object features guide information processing in recurrent neural networks trained for object categorization. arXiv preprint arXiv:2111.07898.

We thank the reviewer for their thoughtful response. We also appreciate the acknowledgment of the potentially conflicting results our study highlights in relation to prior works. Below, we respond to individual questions the reviewer raised:

• Presentation styles We significantly revised the text in section 4.3 to improve its clarity as the reviewer suggested. Since updating the submission is not allowed during this period, we provide a brief summary of changes: 1) expanding the definition of hypotheses 1-3, adding references to prior related work to each hypothesis; 2) functional interpretation of each hypothesis (Hypothesis 1: sustained WM; Hypothesis 2: dynamic updating; Hypothesis 3: dynamic updating of stimulus-specific memory spaces); 2) adding more references to relevant figures containing definitions; 3) further methodological clarification (time steps and labels used for fitting and testing each decoder); 4) clarified the goal of several analyses (e.g., why did we use the Procrustes analysis, and how was it implemented); 5) added geometric interpretation for each result. We believe that these changes have substantially improved the readability and clarity of the results presented in section 4.3 and hope that will address the reviewer’s concern.

• Mathematical definition of latent space The latent space of the RNN refers to a D-dimensional space $\mathbb{R}^D$ where $D$ is the number of units in the RNN model. Consider a set of $N$ vectors $\{\mathbf{v}_1, \mathbf{v}_2, \ldots, \mathbf{v}_N\}$ in D-dimensional space $\mathbb{R}^D$, that represent $N$ decoders in that space. These vectors span a subspace of $\mathbb{R}^D$ that is the set of all possible linear combinations of these vectors. Mathematically, the subspace $S$ spanned by these vectors ( $\text{span} \{ \mathbf{v}_1, \mathbf{v}_2, \ldots, \mathbf{v}_N \}$) is defined as: $S = \left\{ \sum_{i=1}^N \alpha_i \mathbf{v}_i \mid \alpha_i \in \mathbb{R} \right\}$. We will revise the text to make this information and the relation between the decoder weights and the RNN subspace more explicit.

• Relevant citations We thank the reviewer for pointing us to these papers. We agree that these two references are highly relevant and we will add them to the updated manuscript and adjust the text accordingly. We also identify the following references relevant to our studies.

• Questions regarding cross-time generalization analysis We like to first clarify that in our experiments, each model executes exactly one step of computation per input observation and in that sense is different from many prior RNN models used in the literature that involve hundreds of steps within each trial to investigate attractor dynamics in high temporal resolution. For that reason, our experiments do not face the issue of transient responses between two consecutive inputs. In our experiments, we tested the generalization of the decoders up to five steps (the duration of the trial) after observing a stimulus. Thus, if any transient response fades during the trial, our analysis would be able to capture that phenomenon.

References

1. Kozachkov, Leo, John Tauber, Mikael Lundqvist, Scott L. Brincat, Jean-Jacques Slotine, and Earl K. Miller. “Robust and Brain-like Working Memory through Short-Term Synaptic Plasticity.” PLOS Computational Biology 18, no. 12 (December 27, 2022): e1010776.
2. Curtis, Clayton E., and Thomas C. Sprague. “Persistent Activity During Working Memory From Front to Back.” Frontiers in Neural Circuits 15 (2021).
3. Mejías, Jorge F, and Xiao-Jing Wang. “Mechanisms of Distributed Working Memory in a Large-Scale Network of Macaque Neocortex.” Edited by Tatiana Pasternak and Tirin Moore. eLife 11 (February 24, 2022): e72136.
4. Murray, John D., Alberto Bernacchia, Nicholas A. Roy, Christos Constantinidis, Ranulfo Romo, and Xiao-Jing Wang. “Stable Population Coding for Working Memory Coexists with Heterogeneous Neural Dynamics in Prefrontal Cortex.” Proceedings of the National Academy of Sciences 114, no. 2 (January 10, 2017): 394–99.

We understand the reviewer's continuing concern in terms of the clarity of the revised text given that updates to the manuscript are not allowed at this stage. To help the reviewer with their judgement, below we include all of the text from the revised section 4.3, hoping that the reviewer can appreciate the improvements made to the clarity of the new revised section. We highlighted the parts of the text with substantial changes.

Section 4.3

Having examined the encoding of objects in RNN latent space, we next investigated how RNN dynamics enable simultaneous encoding, maintenance, and retrieval of information. Performing our N-back task suite required the RNN to keep track of prior objects’ properties as well as the incoming stimuli with minimal interference. We reasoned that the RNN may implement one of three possible mechanisms to perform the n-back working memory task suite (Figure 4e).

• H1: Slot-based memory subspaces (Luck and Vogel 1997, Whittington et al. 2023). Where the RNN latent space is divided into separate subspaces that are indexed by time within the sequence. Each object is encoded into its corresponding subspace (i.e. slot) and is maintained there until retrieved. By definition, the subspace assigned to each memory slot is distinct and “sustained” in time.

• H2: Relative chronological memory subspaces. Where the RNN latent space is divided into separate subspaces that each maintains object information according to their age (i.e. how long ago they were observed, for example memory of previous observation or prior to last observation). Such a mechanism will require a dynamic process for updating the content of each memory space at each time step during the task.

• H3: Stimulus-specific relative chronological memory subspaces. Which is similar to the relative chronological memory hypothesis but with independent subspaces assigned to each object. Each observation in the sequence is thus encoded into a distinct subspace and encoding of each stimulus is in turn distinctly transformed into associated memory representations.

To identify the hypothesis that best matches the computations performed by the RNN, we analyzed how the RNN latent subspaces that encodes each observed object property (E(S,T ) ) is transformed across time into memory representations (M(S,T ) )(Fig. 1d).

We first tested whether object information is maintained in a temporally stable subspace (i.e. sustained working memory representation) which aligns with H1 prediction (i.e. E(S=i,T =1) =? M(S=i,T =k), k ∈ {2, 3, 4...}; Fig. 4a). For this, we trained decoders to predict the value of each object property using the RNN unit activity during the encoding phase (i.e. Encoding Space) and evaluated its generalization performance in consecutive steps (i.e. Memory Spaces). We reasoned that if the object information is encoded in a subspace that is stable across time (as in a memory slot), the decoders’ generalization performance should be high and comparable to its performance during the encoding phase. Contrary to H1’s prediction, we found that the decoders do not generalize well (Figure 4b), suggesting that the object information is not stably encoded in a temporally-fixed RNN latent space. However, we observed that unlike STSF models, in STMF and MTMF models, the decoding accuracy at the step where the object information needs to be recalled for comparison with the most recent object (i.e. the executive step) is consistently higher than other time steps (Figure 4b and c). This suggests that the object representation is partially realigned with its original encoding representation at the step where a comparison is to be made (Fig. 4b; slightly higher decoding accuracy at the executive step).

Next, we examined whether the object Encoding Space is shared between all incoming stimuli within the sequence (H2) or not (H3) (i.e. E(S=i,T =i) =? E(S=j,T =j)) – the primary difference between H2 and H3 hypotheses. We thus fitted classifiers to decode each object property using the hidden activity from encoding phase (i.e. Encoding Space) of each stimulus within the sequence (i.e. decoding S=i from E(S=i,T =i)), testing it on the stimuli appearing at other time steps (i.e. decoding S=j from E(S=j,T =j)). We performed this analysis for all object properties and for all model operating modes (i.e. each individual task). The validation and generalization accuracies were almost identical (Figure 4d), suggesting a stable encoding representation (E(S=i,T =i)=E(S=j,T =j)) consistent with H2. In other words, each object in the sequence was encoded within the same RNN latent subspace regardless of its order in the sequence.

Continued in the next comment

Dear Authors,

I understand your desire and ambition to continue the discussion to increase my current score. However, likely unknowingly, you are utilizing a rebuttal strategy known as "wearing down the reviewers." Such a strategy might work for less experienced reviewers, but I simply used it to update my priors with more evidence for a bad behavior I suspected (more on this later). In my latest response, I made it crystal clear that 6 was the highest score I am comfortable giving, the best response to this would be to thank me for my time and wish me well as well.

Since you want to continue, let me explain why I cannot go beyond 6. The initial submission feels incomplete, as if it is the version that you could get done until the deadline (also the assigned number of 20855 suggests it was one of the last submitted papers).

To some extent, I assigned a solid chance for it being a placeholder for the extensive revisions you planned to perform after the deadline has passed. As several reviewers noted, methodology is incomplete, results are rushed, and several definitions missing. It is inappropriate to submit a half-finished work to a conference, in which reviewers are already overburdened. It is even more inappropriate to wear those reviewers down with extensive revisions. Thus, though this work was initially appropriate for a poster thanks to the novel science, the possibility of it being a placeholder prevented me from giving a score higher than 6 for a spotlight recommendation.

Now, with your latest response that seems to add major revisions to your work, my priors are updated and I now more strongly suspect that this is indeed what happened. I do not believe this type of behavior should be rewarded, but I will leave it to the AC to decide if such major revisions would be appropriate for the NeurIPS conference.

Due to the reasons described above and my increased suspicion, I will decrease my score down to 5. To be perfectly clear and candid, my score can only go down from here, not up. I hope you will take this as a learning experience for future and once again wish you all the best.

We thank the reviewer’s recognition of the strengths of our manuscript, particularly noting the principled and rigorous nature of the approach. We are glad that the extensive training across various model-task permutations robustly supported our conclusions. In response to the weaknesses and limitations the reviewer highlighted, we carefully provide detailed responses addressing each point:

• Lack of control experiments and whether naturalistic stimuli can be replaced by random vectors. We used a pretrained CNN (on ImageNet), which was trained to categorize objects. This suggests that the output layer of the CNN is, overall, lower-dimensional than its inputs (given that the number of objects classified is fewer than the number of total samples in ImageNet). Nevertheless, to verify this intuition on images sampled from our dataset, we measured the dimensionality of the inputs relative to the dimensionality of the CNN output layer, and the dimensionality of the RNN space. As the Reviewer suggested, we also used randomly sampled vectors (the same number as the number of images in our dataset), and measured the dimensionality of the stimulus vectors in the input space, CNN output layer, and RNN layer. We then compared the dimensionality of our stimulus set vs. randomly sampled vectors for each of these embeddings. We found that the dimensionality of randomly sampled vectors is significantly higher than the dimensionality of the stimuli used in our experiments, suggesting that our results are unique to our specific chosen stimuli and task, and cannot be replaced by using randomly sampled vectors (see fig 1a in the attached 1-page PDF). Importantly, to directly address the Reviewer’s comment, what this control analysis shows is that the output of the final layer of the pretrained CNN is different from being a random vector of numbers, since the dimensionality of random vectors is higher-dimensional than input vectors from our stimulus set. Since dimensionality measures the orthogonality between pairs of vectors (stimuli), our analysis shows that the degree of orthogonalization differs (and cannot be directly reproduced) with random vectors.
• Lack of details in the Methods. We thank the reviewer for bringing this point to our attention. All models were trained and tested on trials of the same length (6 time steps). Moreover, we have included additional methodological (e.g., training details) in our general response to all reviewers. In the revised version of our manuscript, we include those details, in addition to other missing method details.
• Limitations in the discussion. Thank you for bringing this issue to our attention. We have substantially revised the Discussion and Conclusion sections of the paper, expanding on the current study’s limitations and likely future directions. Since updating the submission is not allowed during this period, we provide a brief summary as follows: 1) our findings are limited to the N-back task structure and it’s unclear whether similar computational strategies will emerge in other tasks requiring working memory. However, we provide specific predictions and further experiments that may lead to the discovery of more general principles used by RNNs for different WM tasks (e.g., in a simple delayed-match-to-sample task with only a fixation for the delay, we hypothesize that there will not be a rotation of representations through time); 2) we haven’t considered novel objects in our analyses. We reported reduced performance for novel objects but did not examine the representational geometry with those stimuli or the possible reasons for the diminished performance. We welcome any further suggestions from the reviewer; 3) Our findings are limited to the commonly-used RNN architectures, and we cannot make definitive claims about the representational geometry resulting from other neural network architectural choices. Additionally, we have not examined how scaling the network affects the results; 4) Our results suggest a possible hypothesis about working memory, but further investigation with real neural data is needed to validate this. It would also be valuable to explore the relationship between architectural choices and experimental outcomes.

It sounds like there are 192 $\times$ 4 different stimuli instead of 192? I will just arbitrarily assume its 192 in my argument below but replace it with the correct number if I'm wrong.

I would normally consider what the authors did as a strawman argument. Noise aside, 192 stimuli identities would have a maximum of 192 dimensions. Comparing them to 5000 randomly generated vectors makes absolutely no sense at all. But I would like to give the authors the benefit of the doubt and clarify myself:

Let $x$ be the final layer of the CNN. $W_{xh} x$ is the input to the RNN, where $W_{xh}$ is the set of weights that transforms CNN outputs into RNN inputs. I am saying if you replace $x$ with 192 random vectors, each representing one identity of your stimuli, and $W_{xh}$ is trainable, it is possible to get the same results. Nothing before $W_{xh} x$ matters.

The title of the paper contains the phrase "naturalistic object representations in recurrent neural network models", and I am challenging the very fact that the input to the RNN is not naturalistic. In fact, prior literature has shown that WM orthogonalization can occur with random inputs [1]. It is extremely important that the authors distinguish their model from a model that does not have a CNN, yet capable of producing the same phenomenon.

[1] https://doi.org/10.1371/journal.pcbi.1011555