Dear Mrs. Sörensen,

Thank you very much for submitting your manuscript "Mechanisms of human dynamic object recognition revealed by sequential deep neural networks" for consideration at PLOS Computational Biology. As with all papers reviewed by the journal, your manuscript was reviewed by members of the editorial board and by several independent reviewers. The reviewers appreciated the attention to an important topic. Based on the reviews, we are likely to accept this manuscript for publication, providing that you modify the manuscript according to the review recommendations. Please clarify the availability of any data used in your manuscript. 

Please prepare and submit your revised manuscript within 30 days. If you anticipate any delay, please let us know the expected resubmission date by replying to this email.

When you are ready to resubmit, please upload the following:

[1] A letter containing a detailed list of your responses to all review comments, and a description of the changes you have made in the manuscript. Please note while forming your response, if your article is accepted, you may have the opportunity to make the peer review history publicly available. The record will include editor decision letters (with reviews) and your responses to reviewer comments. If eligible, we will contact you to opt in or out

[2] Two versions of the revised manuscript: one with either highlights or tracked changes denoting where the text has been changed; the other a clean version (uploaded as the manuscript file).

Important additional instructions are given below your reviewer comments.

Thank you again for your submission to our journal. We hope that our editorial process has been constructive so far, and we welcome your feedback at any time. Please don't hesitate to contact us if you have any questions or comments.

Sincerely,

Matthieu Louis

Academic Editor

PLOS Computational Biology

Marieke van Vugt

Section Editor

PLOS Computational Biology

***********************

A link appears below if there are any accompanying review attachments. If you believe any reviews to be missing, please contact ploscompbiol@plos.org immediately:

Reviewer's Responses to Questions

Comments to the Authors:

Please note here if the review is uploaded as an attachment.

Reviewer #1: The authors present a thorough analysis and potential explanations for how the human visual system processes dynamic input, a topic that is currently receiving much attention. The authors made their data and models available which allowed me to reproduce the results and see the code. This is an excellent paper, well written and very interesting, and I only have minor suggestions.

When introducing the BLnet on p5, the authors should justify the rationale for choosing BLnet over any other DNNs.

In the next paragraph, BLnext is mentioned but it has not been explained yet what BLnet is and how it is a modification of BLnet

In Table 1, one may wonder why there is no option for a model that uses dynamic input but without lateral recurrence

P4 adaption->adaptation

Reviewer #2: In this paper, Sörensen and colleagues study dynamic object recognition in recurrent convolutional neural networks, some of which were augmented with neural adaptation mechanisms. Feedforward convolutional neural networks (CNN) have become powerful tools to model ventral stream processing, but they lack the temporal dynamics of our visual system. Previous recurrent CNN models have temporal dynamics, but those studies focused on the “depth through recurrence” principle for static input. In contrast, Sörensen and colleagues studied this class of models with more natural, dynamic input. They show that, like in humans, recurrence hinders processing under highly dynamic conditions, but that some of that impediment is rescued by a neuronally intrinsic adaptation mechanism known to be prevalent in sensory neurons.

I thought this was a carefully conducted computational study that touches upon many subdomains of the field by bridging the gap between neural mechanisms and complex behavior. I greatly appreciate how the authors combined dynamic processes from previous models, and challenged the resulting model with a new task in order to further our understanding of the role of these processes. This led not only to an improved model of dynamic ventral stream processing, but also to a novel perspective on adaptation, namely its potential role in modulating _recurrent_ processing under dynamic conditions. This work neatly shows how complex, task-performing neural networks are an important tool to shed light on neural mechanisms and how they might interact. Specifically for adaptation, up until now the functional significance of this mechanism has been subject to mostly speculation.

Overall I felt like the paper was well written and clear, and the conclusions are well supported. I do have some comments that I would like to see addressed or that need clarification for me. I will discuss these in more detail below.

MAJOR

One major question that I had which was left unaddressed, is how adaptation affects the performance of a recurrent CNN without preceding stimulus. That is, if only one stimulus is presented (e.g. like in the regime where BLnet was trained, or if the target is the first stimulus of a sequence), does adaptation: (a) negatively impact performance, (b) have no real effect, or (c) help?

Answering this question would help to understand whether adaptation specifically helps in a fast-changing environment, or whether it just generally impairs recurrence. If it (a) generally impairs recurrence, adaptation may be helpful in a dynamic context, but could be hurtful when the recurrence is needed in a more static context. Alternatively, (b) it could be that adaptation helps specifically in a dynamic context because it prioritizes feedforward input over recurrence only after the input has changed drastically. In that case there might be no real effect in situations where the input is more constant. A third possibility (c) is that adaptation always renders recurrence more efficient – also when there is no preceding stimulus – by for example dampening redundant feedforward signals.

MINOR

- I was curious how a feedforward model with adaptation would perform. Adaptation by itself could increase performance by highlighting target-specific information, but it could also have a detrimental effect.

- Abstract: What result does this sentence refer to (specifically: it sounds like it refers to the same results as the previous sentence): “Importantly, sequential lateral-recurrent integration was also necessary for a model to display a plausible temporal correspondence with human performance”.

- Fig 2, legend: “representational strength can be computed … as a distance” -> I presume not distance, but similarity/correlation?

- P10: “We evaluated all sequential models on 500 randomly generated images sequences” -> It would be useful to specify whether the target could be placed at all 6 positions.

- Fig 5B: Is this for all stimulus durations or for 80 ms only?

- P20, Methods, Categorical prototypes: What is a “foil” image?

- P20, Methods, Categorical prototypes: “These images stem from the 2-AFC task…” -> does “these images” refer to the foil images of the previous sentence?

- P24, Methods, Single-trial prediction, Predictivity: I was confused by the equation for “RR”. I thought RR was calculated for each specific trial, but a target cannot be both absent and present in the same trial. In other words, if a target is present, FP and TN don't apply, and if a target is absent TP and FN don't apply. Or am I missing something?

Reviewer #3: The article contrasts human performance in RSVP tasks (with rapidly changing images) with deep neural networks trained on image recognition. While humans show performance decrements when the rate of image presentation increases, feedforward deep nets are insensitive to this manipulation by construction. The authors ask if a recurrent type of convolutional network (BLnext, an adaptation of the BLnet model from Spoerer et al) would show similar behavior as humans. And indeed, the recurrent network is affected by faster presentation speeds (less steps per image). Considering that BLnet was trained for optimal recognition over 8 time steps with a constant input image, this finding is not altogether surprising (the less time steps are allowed for each image, the more the testing conditions will depart from the training). Nonetheless, there is some amount of quantitative correspondence between the performance changes observed in humans and in the model. In addition, trial-by-trial variations in accuracy are also matched between humans (at a particular RSVP presentation speed) and the BLnext model at a fixed number of time steps per image, and this match is higher than for feedforward networks. This latter result is non-trivial in my opinion. Finally, the authors also investigate the potential effects of neural adaptation on the performance time course, and find approximately the same pattern of results, but with faster dynamics. That is, the number of time steps that best corresponds to 1 second of processing in the human visual system is lower for a recurrent model with adaptation than without. Although this is not direct proof, the authors argue (based on parsimony) that this is compatible with a computational role for adaptation in explaining the temporal dynamics of human vision.

The paper is well written, the methods appear solid and the results reliable. Overall, the authors have spun their results into a plausible story, which could make a meaningful contribution to the literature on matching human behavior and deep nets. My main comments will thus mirror my lukewarm assessment above.

1. About the main finding, that a pretrained recurrent model is affected by changing image presentation rate (just like humans), while a feedforward network isn’t, nor is a version of the recurrent model that is artificially reset between image changes: I honestly don’t know what else could have been expected here. I agree that (as far as I know) this hasn’t been shown before, and that it is therefore worth reporting. But on the other hand, I don’t see an alternative outcome that could have made sense.

2. Related to this, it could have been informative to compare different types of recurrent networks, with different connectivity patterns. For instance, Spoerer et al (2017) had contrasted the BL architecture (lateral connections) with a BT architecture (top-down connections) and a BLT version (both connection types). I think all three types of architectures, if trained on a constant input image, would be affected by changing the input to an RSVP stream. But it’s not clear to me which of the three would best match human performance, so this could have been a more interesting question. As far as I know, there is no pretrained version of BT and BLT on the ecoset dataset that could be compared against BLnet, so I understand why authors did not do this, but it could be a worthwhile consideration for future work. It certainly would have made the paper more interesting for me.

**********

Have the authors made all data and (if applicable) computational code underlying the findings in their manuscript fully available?

The PLOS Data policy requires authors to make all data and code underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data and code should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data or code —e.g. participant privacy or use of data from a third party—those must be specified.

Reviewer #1: Yes

Reviewer #2: None

Reviewer #3: No: I didn't see a code availability statement

**********

PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files.

If you choose “no”, your identity will remain anonymous but your review may still be made public.

Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy.

Reviewer #1: Yes: Tijl Grootswagers

Reviewer #2: No

Reviewer #3: No

Figure Files:

While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, https://pacev2.apexcovantage.com. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email us at figures@plos.org.

Data Requirements:

Please note that, as a condition of publication, PLOS' data policy requires that you make available all data used to draw the conclusions outlined in your manuscript. Data must be deposited in an appropriate repository, included within the body of the manuscript, or uploaded as supporting information. This includes all numerical values that were used to generate graphs, histograms etc.. For an example in PLOS Biology see here: http://www.plosbiology.org/article/info%3Adoi%2F10.1371%2Fjournal.pbio.1001908#s5.

Reproducibility:

To enhance the reproducibility of your results, we recommend that you deposit your laboratory protocols in protocols.io, where a protocol can be assigned its own identifier (DOI) such that it can be cited independently in the future. Additionally, PLOS ONE offers an option to publish peer-reviewed clinical study protocols. Read more information on sharing protocols at https://plos.org/protocols?utm_medium=editorial-email&utm_source=authorletters&utm_campaign=protocols

References:

Review your reference list to ensure that it is complete and correct. If you have cited papers that have been retracted, please include the rationale for doing so in the manuscript text, or remove these references and replace them with relevant current references. Any changes to the reference list should be mentioned in the rebuttal letter that accompanies your revised manuscript.

If you need to cite a retracted article, indicate the article’s retracted status in the References list and also include a citation and full reference for the retraction notice.

Dear Mrs. Sörensen,

We are pleased to inform you that your manuscript 'Mechanisms of human dynamic object recognition revealed by sequential deep neural networks' has been provisionally accepted for publication in PLOS Computational Biology.

Before your manuscript can be formally accepted you will need to complete some formatting changes, which you will receive in a follow up email. A member of our team will be in touch with a set of requests.

Please note that your manuscript will not be scheduled for publication until you have made the required changes, so a swift response is appreciated.

IMPORTANT: The editorial review process is now complete. PLOS will only permit corrections to spelling, formatting or significant scientific errors from this point onwards. Requests for major changes, or any which affect the scientific understanding of your work, will cause delays to the publication date of your manuscript.

Should you, your institution's press office or the journal office choose to press release your paper, you will automatically be opted out of early publication. We ask that you notify us now if you or your institution is planning to press release the article. All press must be co-ordinated with PLOS.

Thank you again for supporting Open Access publishing; we are looking forward to publishing your work in PLOS Computational Biology. 

Best regards,

Matthieu Louis

Academic Editor

PLOS Computational Biology

Marieke van Vugt

Section Editor

PLOS Computational Biology

***********************************************************

Reviewer's Responses to Questions

Comments to the Authors:

Please note here if the review is uploaded as an attachment.

Reviewer #1: The authors have addressed all my comments and improved the manuscript significantly in this revision.

Reviewer #2: Sörensen and colleagues fully addressed my questions and I have no remaining concerns.

I would like to add one suggestion, which is that the authors could expand a bit on the implications of the new analysis of Fig. S2 in the manuscript (either briefly in the results, or in the discussion). From the short description currently in the results I am not sure whether the implications of the analysis are as clear to the reader as they are in the author’s response to me, which was:

“Fig. S2 shows that whereas BLnet models with adaptation are indistinguishable from those without adaptation for a few model steps per image, adaptation models show degraded performance when the same image is processed longer (i.e., > 4 model steps). Yet, notably, all models (incl. those with adaptation) still show comparable maximum performance. These results fit with our account that the effects of adaptation become most apparent once the benefits of lateral recurrent processing have been reaped, that is, once a model’s recurrent processing becomes repetitive. The result in Fig. S2 also clarifies how including adaptation during sequential processing helps recognition: Once a lateral recurrent model has reached its peak performance, adaptation helps to ready a model for processing new incoming stimuli. Together, these results provide further support for the notion that lateral recurrence and adaptation mutually enhance each other.”

Reviewer #3: My initial review stated that the findings, although unsurprising, could be considered worth reporting.

This has not changed for the revised manuscript.

**********

Have the authors made all data and (if applicable) computational code underlying the findings in their manuscript fully available?

The PLOS Data policy requires authors to make all data and code underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data and code should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data or code —e.g. participant privacy or use of data from a third party—those must be specified.

Reviewer #1: None

Reviewer #2: None

Reviewer #3: None

**********

PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files.

If you choose “no”, your identity will remain anonymous but your review may still be made public.

Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy.

Reviewer #1: No

Reviewer #2: No

Reviewer #3: No