We thank the reviewer for their positive assessment of our work and their helpful comment.

We agree that the insights provided by our observations were insufficiently clear. The primary contributions of this study are methodological rather than hypothesis-driven. We have now amended the manuscript to indicate:

“Our approach provides three main contributions: our MEG decoder (1) yields a 7X increase in performance as compared to linear baselines (Fig. 2), (2) helps reveal when high-level semantic features are processed in the brain (Fig. 3) and (3) allows the continuous generation of images from temporally-resolved brain signals (Fig. 4). Overall, this approach thus paves the way to better understand the unfolding of the brain responses to visual inputs.”

In addition, to clarify and substantiate the insights provided by this study, we have now added several new analyses.

First, we now implement a high-temporally resolved analysis of retrieval and generative decoders, using both sliding and growing windows (Updated figures 3, 4; new figures S6A, S7, S9 ). Overall, these results clarify when low- and high-level representations can be decoded from brain activity, and thus shed lights on the dynamics of visual processing in the brain:

“Consistent with the decoding peaks observed after image onset and offset (Fig. 3), the retrieval performance of all growing-window models considerably improves after the offset of the image. Together, these results suggest that the brain activity represents both low- and high-level features even after image offset. This finding clarifies mixed results previously reported in the literature. Carlson et al., (2011; 2013), reported small but significant decoding performances after image offset. However, other studies (Cichy et al., 2014, Hebart et al., 2023) did not observe such a phenomenon. In all of these cases, decoders were based on pairwise classification of object categories and on linear classifiers. The improved sensitivity brought by (1) our deep learning architecture, (2) its retrieval objective and (3) its use of pretrained latent features may thus help clarify the dynamics of visual representations in particular at image offset. We speculate that such offset responses could reflect an intricate interplay between low- and high-level processes that may be difficult to detect with a pairwise linear classifier. Overall, we hope that the present methodological contribution will help shine light on this understudied phenomenon.”

References:

Carlson, Thomas A., et al. "High temporal resolution decoding of object position and category." Journal of vision 11.10 (2011): 9-9.
Carlson, Thomas, et al. "Representational dynamics of object vision: the first 1000 ms." Journal of vision 13.10 (2013): 1-1.
Cichy, Radoslaw Martin, Dimitrios Pantazis, and Aude Oliva. "Resolving human object recognition in space and time." Nature neuroscience 17.3 (2014): 455-462.
Hebart, Martin N., et al. "THINGS-data, a multimodal collection of large-scale datasets for investigating object representations in human brain and behavior." Elife 12 (2023): e82580.

Second, we inspect the model with a new analysis of its weights (Appendix A.10) and clarify that the decoding time course is different for low- and high-level features:

“We inspect our decoders to better understand how they use information in the time domain. To do so, we leverage the fact that our architecture preserves the temporal dimension of the input up until the output of its convolutional blocks. This output is then reduced by an affine transformation learned by the temporal aggregation layer (see Section 2.3 and Appendix A.1). Consequently, the weights $w^{agg} \in \mathbb{R}^T$ can reveal on which time steps the models learned to focus. To facilitate inspection, we initialize $w^{agg}$ to zeros before training and plot the mean absolute weights of each model (averaged across seeds).

The results are presented in Fig. S8. While these weights are close to zero before stimulus onset, they deviate from this baseline after stimulus onset, during the maintenance period and after stimulus offset. Interestingly, and unlike high-level features (e.g., VGG-19, CLIP-Vision), low-level features (e.g., color histogram, AutoKL and DINOv2) have close-to-zero weights in the 0.2-0.5 s interval.

This result suggests that low-level representations quickly fade away at that moment. Overall, this analysis demonstrates that the models rely on these three time periods to maximize decoding performance, including the early low-level responses ($t=0-0.1$s).“

1.B. New analyses

To highlight the scientific insights that our MEG decoding may bring, we implemented several new analyses and updated: Figure 3, which now shows retrieval performance at a much higher temporal resolution; Figure 4, which now shows image generations as a function of time; Supplementary figures (growing window retrieval, model weights analysis and time-resolved image generation metrics in Appendices 8,10 and 11, respectively), which detail when low- and high-level features may be decoded from brain activity.

Overall, these new analyses reveal that (1) late brain responses specifically represent high-level features and (2) the brain responses to image offset represent both low and high-level responses. These elements subsume previous neuroscientific investigation. We now indicate:

In Appendix 5:

“Consistent with the decoding peaks observed after image onset and offset (Fig. 3), the retrieval performance of all growing-window models considerably improves after the offset of the image. Together, these results suggest that the brain activity represents both low- and high-level features even after image offset. This finding clarifies mixed results previously reported in the literature. Carlson et al., (2011; 2013), reported small but significant decoding performances after image offset. However, other studies (Cichy et al., 2014, Hebart et al., 2023) did not observe such a phenomenon. In all these cases, decoders were based on pairwise classification of object categories and on linear classifiers. The improved sensitivity brought by (1) our deep learning architecture, (2) its retrieval objective and (3) its use of pretrained latent features may thus help clarify the dynamics of visual representations in particular at image offset. We speculate that such offset responses could reflect an intricate interplay between low- and high-level processes that may be difficult to detect with a pairwise linear classifier. We hope that the present methodological contribution will help shine light on this understudied phenomenon.”

References:

Carlson, Thomas, et al. "Representational dynamics of object vision: the first 1000 ms." Journal of vision 13.10 (2013): 1-1.
Cichy, Radoslaw Martin, Dimitrios Pantazis, and Aude Oliva. "Resolving human object recognition in space and time." Nature neuroscience 17.3 (2014): 455-462.
Hebart, Martin N., et al. "THINGS-data, a multimodal collection of large-scale datasets for investigating object representations in human brain and behavior." Elife 12 (2023): e82580.

In Appendix 10:

“We inspect our decoders to better understand how they use information in the time domain. To do so, we leverage the fact that our architecture preserves the temporal dimension of the input up until the output of its convolutional blocks. This output is then reduced by an affine transformation learned by the temporal aggregation layer (see Section 2.3 and Appendix A.2). Consequently, the weights $w^{agg} \in \mathbb{R}^T$ can reveal on which time steps the models learned to focus. To facilitate inspection, we initialize $w^{agg}$ to zeros before training and plot the mean absolute weights of each model (averaged across seeds).

The results are presented in Fig. S8. While these weights are close to zero before stimulus onset, they deviate from this baseline after stimulus onset, during the maintenance period and after stimulus offset. Interestingly, and unlike high-level features (e.g., VGG-19, CLIP-Vision), low-level features (e.g., color histogram, AutoKL and DINOv2) have close-to-zero weights in the 0.2-0.5 s interval.

This result suggests that low-level representations quickly fade away at that moment. Overall, this analysis demonstrates that the models rely on these three time periods to maximize decoding performance, including the early low-level responses ($t=0-0.1$s).“

Finally, we now provide a time-resolved analysis of low- and high-level reconstruction metrics (Appendix A.11). These results confirm that the early responses to both image onset and offset are primarily associated with low-level metrics. On the other hand, the decoding of brain responses during the second half of the image presentation interval (0.2-0.5 s) appears to be more related to high-level features.

We thank the reviewer for the thorough review and insightful suggestions that effectively help improve the quality of our manuscript.

1. Weakness

We agree with the reviewer that the scientific insight provided by our study could have been clearer. We have now (1) updated the contribution section, (2) provided new analyses and (3) amended the discussion section.

1.A. Contribution

This study primarily focuses on methodological contributions. We have now amended the manuscript as follows:

“Our approach provides three main contributions: our MEG decoder (1) yields a 7X increase in performance as compared to linear baselines (Figure 2), (2) helps reveal when high-level semantic features are processed in the brain (Figure 3) and (3) allows the continuous generation of images from temporally-resolved brain signals (Figure 4). Overall, this approach thus paves the way to better understand the unfolding of the brain responses to visual inputs.“

I've increased my score to 8, I am more convinced of the scientific contribution.

I still want to comment on the sub-optimal presentation of the rebuttal itself.

While the authors have indeed provided additional scientific justification which I find convincing, the authors have not addressed my requests for clarification, and the rebuttal is not well formatted.

The revision PDF also does not include highlights as typical in ICLR rebuttals to visualize the changes.

Hopefully the authors can improve upon this in the future.

Edit: Reading the additional clarifications below, the authors use the last layer (1000 class output) of the each network as the baseline, which does not strike me as particularly sound. In this layer, there exists an exact optima, where it is an indicator function of the correct class. It is not clear that this output is necessarily reflective of different architectures. And to my knowledge, using a linear probe on the last categorical layer is not done in other works. I'm re-calibrating my score to a 6.

1.C. Updated impact

Finally, we modified the manuscript to further discuss what scientific insights our approach may provide. The present work is primarily a methodological contribution, and a demonstration that natural images can be decoded from MEG at a much better level than anticipated (7X as compared to linear baseline). We updated the discussion section (paragraph on Impact):

“Our methodological contribution has both fundamental and practical impacts. First, decoding the unfolding of perceptual representations could clarify the unfolding of visual processing in the brain. While there is considerable work on this issue, neural representations are challenging to interpret because they represent latent, abstract, feature spaces. Generative decoding, on the contrary, can provide concrete and, thus, interpretable predictions. Put simply, generating images at each time step could help neuroscientists understand whether specific – potentially unanticipated – textures or object parts are represented. For example, Cheng et al., (2023) showed that generative decoding applied to fMRI can be used to decode the subjective perception of visual illusions. Such techniques can thus help to clarify the neural bases of subjective perception and to dissociate them from those responsible for ``copying’’ sensory inputs. Our work shows that this endeavor could now be applied to clarify when these subjective representations arise.

Second, generative brain decoding has concrete applications. For example, it has been used in conjunction with encoding, to identify stimuli that maximize brain activity (Bashivan et al., 2019). Furthermore, non-invasive brain-computer interfaces (BCI) have been long-awaited by patients with communication challenges related to brain lesions. BCI, however, requires real-time decoding, and thus limits the use of neuroimaging modalities with low temporal resolution such as fMRI. This application direction, however, will likely require extending our work to EEG, which provides similar temporal resolution to MEG, but is typically much more common in clinical settings.“

References
Cheng, Fan L., et al. "Reconstructing visual illusory experiences from human brain activity." Science Advances 9.46 (2023): eadj3906.

We thank our reviewer for their thorough assessment and their helpful comments.

1. Unimpressive reconstruction.

We agree that the quality of generated images are low in comparison to what can be obtained with 7T fMRI (Table 1).

However, we respectfully disagree that this is not “impressive”. To the best of our knowledge, this is the first study showing MEG-based reconstruction of visual processes. Relative to fMRI, MEG has a very low spatial resolution. Consequently, reconstructing these rich contents is far from obvious – for what it’s worth, we certainly did not expect these results.

We thus clarify our contribution as follows:

“Our approach provides three main contributions: our MEG decoder (1) yields a 7X increase in performance as compared to linear baselines (Fig. 2), (2) helps reveal when high-level semantic features are processed in the brain (Fig. 3) and (3) allows the continuous generation of images from temporally-resolved brain signals (Fig. 4). Overall, this approach thus paves the way to better understand the unfolding of the brain responses to visual inputs.”

Beyond these subjective considerations, part of this underwhelming feeling may be due to our original figures. We originally opted to automatically select those with the best, average and worst aggregated score across low-level (SSIM) and high-level metrics (SwAV) as there is no standard metric to rank images decoded from the brain. However, after closer inspection, we realize that this SSIM-SwAV combination may not faithfully reflect whether images are accurately reconstructed. To address this issue, we now added two novel figures (Figures 4 and S6, in Appendix A.7), which show manually-selected cases of successful and failed reconstructions, as well as their dynamics. In particular, we now note:

“Figure S6 shows examples of failed generations. Overall, they appear to encompass different types of failures. Some generations appear to miss the correct category of the true object (e.g., bamboo, batteries, bullets and extinguisher in columns 1-4), but generate images with partially similar textures. Other generations appear to recover some category-level features but generate unrealistic chimeras (bed: weird furniture, alligator: swamp beast; etc, in columns 5-6). Finally, some generations seem to be plain wrong, with little-to-no preservation of low- or high-level features (columns 7-8). We speculate that these different types of failures may be partially resolved with different methods, such as better generation modules (for chimeras) and optimization on both low- and high-level features (for category errors).”

In any case, we agree that these generations are far from being perfect. We thus toned down our abstract by emphasizing the preliminary aspect of our results:

“Overall, these results, while preliminary, provide an important step towards the decoding - in real-time - of the visual processes continuously unfolding within the human brain.”

2. Impact of $\lambda$ on retrieval.

The reviewer is right that we did not include $\lambda$ from Equation 3 in the hyperparameter search presented in Appendix A.2. As such, Figure S1 presents retrieval performance for models trained with the CLIP loss term only (i.e., with $\lambda=1.0$). We decided not to include $\lambda$ in this hyperparameter search because the focus of this search was on the retrieval task, which is well aligned with the CLIP loss. However, as detailed in the next answer, we did implement a hyperparameter selection procedure for $\lambda$, but that is carried out in a second step, as it relates to a generation-based evaluation.

3. $\lambda$ search on test data.

The hyperparameter search for $\lambda$ was performed on the “large” test set. The model obtained with the best $\lambda$ was then used for the generation of images only on the “small” test set. Importantly, the large and the small test sets are disjoint, thus there is no leakage. (Of note, we fixed a typo in the original manuscript: the set of $\lambda$ values should in fact have read {0.0,0.25,0.5,0.75}.)

We now clarify this issue in the methods:

In Section 2.3:

"We select $\lambda$ in $\mathcal{L}_{Combined}$ by sweeping over {0.0,0.25,0.5,0.75} and pick the model whose top-5 accuracy is the highest on the "large" test set (which is disjoint from the "small" test set used for generation experiments; see Section 2.8)."

In Appendix A.2:

"We run a hyperparameter grid search to find an appropriate configuration (MEG preprocessing, optimizer, brain module architecture and CLIP loss) for the MEG-to-image retrieval task."

"We focus the search on the retrieval task, \ie by setting $\lambda=0$ in Eq. 3, and leave the selection of an optimal $\lambda$ to a model-specific sweep using a held-out set (see Section 2.3)."

We thank the reviewer for their positive assessment of our work and their insightful comments. We address every comment and question below and highlight what we changed in the manuscript as a result.

1. We agree that the conceptual novelty is limited: there is a lot of work on the decoding of images from brain activity. However, our method and our results are novel: Past research has focused on either (1) classification (Grootswagers et al., 2019; King & Wyart, 2021) (2) encoding (Cichy et al., 2017; Gifford et al., 2022) tasks from M/EEG or (3) image reconstruction from fMRI (Seeliger et al., 2018; VanRullen & Reddy, 2019; Ozcelik & VanRullen, 2023). To our knowledge, there is no study on image reconstruction from MEG, and thus no model or benchmark to reconstruct natural images from the unfolding of brain activity. We now clarify this issue in the discussion:

“Our approach provides three main contributions: our MEG decoder (1) yields a 7X increase in performance as compared to linear baselines (Fig. 2), (2) helps reveal when high-level semantic features are processed in the brain (Fig. 3) and (3) allows the continuous generation of images from temporally-resolved brain signals (Fig. 4). Overall, this approach thus paves the way to better understand the unfolding of the brain responses to visual inputs.“

To emphasize this point, we updated Figures 3 and 4 and added Appendix A.8, which now show a highly-resolved decoding of MEG responses, based on retrieval and generation with sliding or growing window decoders.

2. We agree that it is interesting to highlight the differences between VGG and other high-level embedding observed after image onset and image offset. This is carried out in Appendix A.11.

“Lastly, we provide a sliding window analysis of these metrics in Appendix A.11. These results suggest that early responses to both image onset and offset are primarily associated with low-level metrics, while high-level features appear more related to brain activity in the 200-500 ms interval."

3. The reviewer is right that we forgot to add uncertainty estimates. We now added SEM (standard error of the mean) for each relevant metric. Finally, we included a new table in Appendix 9 reporting individual metrics and SEM for each participant (Table S3).

We have now uploaded a “diff” of the pdf with changes highlighted in blue in the Supplementary Materials zip file. Of note, the zip file also contains a pdf of the appendices. We would like to thank the reviewers again for their time and feedback.