The present work examined the extent to which task context (i.e. behavioral goals) impact visual object processing. The authors employed EEG (primarily multivariate analysis techniques) to uncover the time windows in which task context affects object processing. The theoretical question posed by the authors was whether task impacts processing during the formation of object representations (i.e. at the early, perceptual stages of visual processing) or after their formation (i.e. at later, post-perceptual stages). Participants performed a verification task on novel objects spanning two orthogonal dimensions (shape and texture). The multivariate analysis results were somewhat mixed. Task effects in the form of a difference in decoding accuracy between task-relevant and task-irrelevant classifiers emerged around 450 ms post-stimulus onset and was evident only in texture- but not in shape- classification. In contrast, cross-task classifiers showed an earlier effect (again, only for the texture task) starting around 250 ms post-stimulus onset. In spite of these differences, the results still suggest that task effects emerge relatively late in time, in line with previous EEG/ERP findings examining the influence of task context on object representations (primarily Hebart et al., 2018). Overall, this is a nicely-written paper presenting an elegant and well-controlled experimental design, featuring two innovative aspects: (1) the use of novel, unfamiliar objects, which minimizes the impact of prior experience or expectations, and (2) utilizing two sets of multivariate analyses examining not only how task impacts the amount of information in object representations, but also the source of this information. The work will definitely add to recent works utilizing E/MEG to reveal how and when cognitive high-level factors impact perceptual processing. However, I have substantial concerns about some of the findings, which require additional elaboration and testing on behalf of the authors. I detail my comments on these concerns and other various points requiring the authors’ further consideration below.

Major comments
(1) My main comment concerns the validity of the central conclusion of the study, namely, that task modulates object representations, given the somewhat inconclusive results of the multivariate analyses. A brief view of both figures 6 and 7 reveals that task effects, either as in the form of task relevance or in the form of cross-task decoding, are observed only in one of the two tasks used in the study (texture classification), meaning that half of the time, task does NOT impact the nature of object processing (in the form of shape classification). In fact, one may even argue that the latter task is more pertinent to object recognition than the former (since form and curvature are more diagnostic for discriminating objects compared to texture), which makes the proposition that task impacts object processing even harder to claim. How do the authors account for such a disparity between these two dimensions? In my view, without demonstrating clear task effects for both tasks, there are no grounds for making any strong conclusions either for or against the role of task context in object processing. Moreover, given the disparity in the results, there is no clear justification for presenting the “overall” time course of decoding, as it is clearly driven by one task but not the other.

(2) One potential reason for the disparity in the results of the decoding analyses for the two tasks might be the behavioral differences between the two tasks. As reported on page 20, participants were significantly slower and less accurate on the texture task compared to the shape task. In other words, the texture task was significantly harder than the shape task for some reason. This difference in task difficulty has two implications for the study. The first one concerns is at the methodological level and concerns the decoding analyses, and the second is at a more general level and concerns the more general interpretations of the current results. Starting with the decoding analyses, previous neuroimaging works have demonstrated that differences in response magnitudes can be easily picked up by multivariate classifiers, and hence may lead to erroneous conclusions regarding the source of the differences between experimental conditions (please see the following works:

Jimura, K., & Poldrack, R. A. (2012). Analyses of regional-average activation and multivoxel pattern information tell complementary stories. Neuropsychologia, 50(4), 544-552.; Smith, A. T., Kosillo, P., & Williams, A. L. (2011). The confounding effect of response amplitude on MVPA performance measures. Neuroimage, 56(2), 525-530.; Davis, T., LaRocque, K. F., Mumford, J. A., Norman, K. A., Wagner, A. D., & Poldrack, R. A. (2014). What do differences between multi-voxel and univariate analysis mean? How subject-, voxel-, and trial-level variance impact fMRI analysis. Neuroimage, 97, 271-283.).

This is especially relevant when one observes differences in task difficulty, which may yield magnitude differences and consequently manifest in decoding differences which are not necessarily reflecting the actual effect of task on object representations. Thus, the existing differences in task difficulty make it hard to assess the current decoding results. To quote Jimura and Poldrak (2012): “…Likewise, any differences between conditions in task difficulty or response time may be sensed by the classifier, but these are potentially downstream effects rather than being causally involved in the decision process.” The second issue with task difficulty is that it poses a general confound to the overall results, irrespective of analysis, as one cannot rule out alternative explanations to current findings. These include a broad gamut of alternative explanations: attention, task-difficulty, response planning, motor execution, etc. Unfortunately, without controlling for task difficulty none of these alternative explanations can be easily brushed away. In sum, the authors should provide more explanation or present more data to ensure the reader that the task effects observed in the univariate and the multivariate analyses do not reflect an artifact stemming from overall differences in task difficulty.

(3) I would have liked to see a little more in-depth discussion of what cross-decoding actually means, and even more importantly, what the difference between task-relevant and cross-task classifiers means at a more theoretical level. For instance, it would be really beneficial if other relevant works are cited, clarifying the notion of the ‘source’ of information in object representations as a way to examine task effects. In that context, I was also missing a statistical comparison of the cross-task classifier performance and the task-irrelevant condition. I believe this contrast is equally important, as it shows unique information above and beyond the task-irrelevant condition. This contrast could coincide with the other two comparisons, and comprise of a single figure with all three traces together.

(4) While I assume the restriction of the analysis to the posterior channels results from the need to enhance the statistical power of the analyses, I would suggest the authors also consider applying similar analyses to the anterior half of the electrodes. The addition of more anterior electrodes, and their analysis in conjunction with the posterior ones may shed light on the spatiotemporal dynamics of top-down interactions in the brain. Assuming such an analysis would yield anterior task effect, it would allow the authors to determine if such effects either precede or follow the posterior task effects. For a similar approach, see Bar et al. (PNAS, 2006) who examined the mechanisms underlying top-down facilitation of object recognition by combining MEG and fMRI. They showed successful object recognition developed in prefrontal regions 50ms earlier than areas along the ventral stream, suggesting that prefrontal regions are involved in the formation of object representations, anticipating activity in visual regions.

(5) The reported ERP findings should be expanded to assess additional components and explore the Task by Object interaction. As such, Figure 5 could be improved by illustrating the Task and Object effects, and their interaction. To do this, the authors could create two difference waveforms comparing the shape dimensions (round minus cubic) in the task relevant (shape) and task irrelevant (texture) tasks. Similarly, they could create difference waveforms for the texture dimension (large minus small) under the task relevant (texture) and task irrelevant (shape) tasks. Creating this 2x2 figure would allow the reader to examine amplitude and latency effects spanning the object type, task type, and interactions. For a similar approach in scene recognition, see Figure 4 in Hansen et al (2018):

Hansen, N. E., Noesen, B. T., Nador, J. D., & Harel, A. (2018). The influence of behavioral relevance on the processing of global scene properties: an ERP study. Neuropsychologia, 114, 168-180.)

(6) Additional discussion may be necessary to discuss the underlying processes involved in the formation of object representations. The advantage of traditional ERP univariate analysis is that it can provide a direction for the effects, which can shed light on the task difficulty issues, as well as differentiating perceptual vs. post-perceptual processes. This discussion can supplement the MVPA findings, which currently fall short of describing the processing dynamics that underlie the formation of object representations.

Minor Comments
(7) Page 7 – the last sentence discusses the penetrability of perception and should cite Pylyshyn’s seminal paper: Pylyshyn, Z. (1999). Is vision continuous with cognition?: The case for cognitive impenetrability of visual perception. Behavioral and brain sciences, 22(3), 341-365.

(8) The levels for Texture described in the paper are “large” and “small”. However, these may not be the most apt descriptors for differentiating this stimulus dimension. Perhaps more appropriate descriptors would be “sparse” and “dense”.

(9) Consider moving Figure 5 closer to whether the EEG results are reported.

I appreciate the substantial work the authors did in revising the manuscript. They have addressed the majority of my comments, and the manuscript is indeed substantially improved. However, there is still an important caveat that in my opinion is not adequately addressed in the revised manuscript. This is of course the caveat of differences in task difficulty (mentioned at length in my original review) and the lack of symmetry in the reported decoding effects between the two tasks. The direction the authors are taking in their selection of six participants with equal task performance is exactly right and points to the anticipated outcome if indeed the level of difficulty was equated. However, at the same time, it highlights that the current results (i.e. for the full sample) are exactly stemming from this lack of control! In other words, it is not sufficient to select a sub-sample of the participants and then report that they show the anticipated effect (also, from a statistical and methodological standpoint, this is definitely not a recommended practice). Thus, I strongly recommend the authors collect a new sample of subjects with (1) a sufficient enough sample size, and (2) equating task difficulty for all tasks, and demonstrate that they can replicate the current findings with the small sub-sample reported. While I understand that this is a timely and cumbersome task, it will guarantee the validity of the claims made by the authors and substantially improve the reliability of the current findings.

I was surprised to see the authors disregarded my former comment/suggestion regarding the crucial need for additional data collection and their restructuring of the paper instead to focus on reporting the results from just a single task. Even without mentioning the methodological pitfalls involved in such cherry-picking, it does not make sense to me how one can discuss the effects of task on object recognition while focusing on one task context, ignoring proper controls of the additional tasks used. Unfortunately, I do not see the merit in the current approach adopted in the present revision and would advise the authors to ensure the validity of their findings by collecting new data with two tasks equated for task difficulty.