Human EEG and artificial neural networks reveal disentangled representations of object real-world size in natural images

Reviewer #1 (Public Review):

Lu & Golomb combined EEG, artificial neural networks, and multivariate pattern analyses to examine how different visual variables are processed in the brain. The conclusions of the paper are mostly well supported, but some aspects of methods and data analysis would benefit from clarification and potential extensions.

The authors find that not only real-world size is represented in the brain (which was known), but both retinal size and real-world depth are represented, at different time points or latencies, which may reflect different stages of processing. Prior work has not been able to answer the question of real-world depth due to the stimuli used. The authors made this possible by assessing real-world depth and testing it with appropriate methodology, accounting for retinal and real-world size. The methodological approach combining behavior, RSA, and ANNs is creative and well thought out to appropriately assess the research questions, and the findings may be very compelling if backed up with some clarifications and further analyses.

The work will be of interest to experimental and computational vision scientists, as well as the broader computational cognitive neuroscience community as the methodology is of interest and the code is or will be made available. The work is important as it is currently not clear what the correspondence between many deep neural network models and the brain is, and this work pushes our knowledge forward on this front. Furthermore, the availability of methods and data will be useful for the scientific community.

Some analyses are incomplete, which would be improved if the authors showed analyses with other layers of the networks and various additional partial correlation analyses.

Clarity

(1) Partial correlations methods incomplete - it is not clear what is being partialled out in each analysis. It is possible to guess sometimes, but it is not entirely clear for each analysis. This is important as it is difficult to assess if the partial correlations are sensible/correct in each case. Also, the Figure 1 caption is short and unclear.

For example, ANN-EEG partial correlations - "Finally, we directly compared the timepoint-by-timepoint EEG neural RDMs and the ANN RDMs (Figure 3F). The early layer representations of both ResNet and CLIP were significantly correlated with early representations in the human brain" What is being partialled out? Figure 3F says partial correlation

Issues / open questions

(2) Semantic representations vs hypothesized (hyp) RDMs (real-world size, etc) - are the representations explained by variables in hyp RDMs or are there semantic representations over and above these? E.g., For ANN correlation with the brain, you could partial out hyp RDMs - and assess whether there is still semantic information left over, or is the variance explained by the hyp RDMs?

(3) Why only early and late layers? I can see how it's clearer to present the EEG results. However, the many layers in these networks are an opportunity - we can see how simple/complex linear/non-linear the transformation is over layers in these models. It would be very interesting and informative to see if the correlations do in fact linearly increase from early to later layers, or if the story is a bit more complex. If not in the main text, then at least in the supplement.

(4) Peak latency analysis - Estimating peaks per ppt is presumably noisy, so it seems important to show how reliable this is. One option is to find the bootstrapped mean latencies per subject.

(5) "Due to our calculations being at the object level, if there were more than one of the same objects in an image, we cropped the most complete one to get a more accurate retinal size. " Did EEG experimenters make sure everyone sat the same distance from the screen? and remain the same distance? This would also affect real-world depth measures.

Reviewer #2 (Public Review):

Summary:

This paper aims to test if neural representations of images of objects in the human brain contain a 'pure' dimension of real-world size that is independent of retinal size or perceived depth. To this end, they apply representational similarity analysis on EEG responses in 10 human subjects to a set of 200 images from a publicly available database (THINGS-EEG2), correlating pairwise distinctions in evoked activity between images with pairwise differences in human ratings of real-world size (from THINGS+). By partialling out correlations with metrics of retinal size and perceived depth from the resulting EEG correlation time courses, the paper claims to identify an independent representation of real-world size starting at 170 ms in the EEG signal. Further comparisons with artificial neural networks and language embeddings lead the authors to claim this correlation reflects a relatively 'high-level' and 'stable' neural representation.

Strengths:

- The paper features insightful figures/illustrations and clear figures.

- The limitations of prior work motivating the current study are clearly explained and seem reasonable (although the rationale for why using 'ecological' stimuli with backgrounds matters when studying real-world size could be made clearer; one could also argue the opposite, that to get a 'pure' representation of the real-world size of an 'object concept', one should actually show objects in isolation).

- The partial correlation analysis convincingly demonstrates how correlations between feature spaces can affect their correlations with EEG responses (and how taking into account these correlations can disentangle them better).

- The RSA analysis and associated statistical methods appear solid.

Weaknesses:

- The claim of methodological novelty is overblown. Comparing image metrics, behavioral measurements, and ANN activations against EEG using RSA is a commonly used approach to study neural object representations. The dataset size (200 test images from THINGS) is not particularly large, and neither is comparing pre-trained DNNs and language models, or using partial correlations.

- The claims also seem too broad given the fairly small set of RDMs that are used here (3 size metrics, 4 ANN layers, 1 Word2Vec RDM): there are many aspects of object processing not studied here, so it's not correct to say this study provides a 'detailed and clear characterization of the object processing process'.

- The paper lacks an analysis demonstrating the validity of the real-world depth measure, which is here computed from the other two metrics by simply dividing them. The rationale and logic of this metric is not clearly explained. Is it intended to reflect the hypothesized egocentric distance to the object in the image if the person had in fact been 'inside' the image? How do we know this is valid? It would be helpful if the authors provided a validation of this metric.

- Given that there is only 1 image/concept here, the factor of real-world size may be confounded with other things, such as semantic category (e.g. buildings vs. tools). While the comparison of the real-world size metric appears to be effectively disentangled from retinal size and (the author's metric of) depth here, there are still many other object properties that are likely correlated with real-world size and therefore will confound identifying a 'pure' representation of real-world size in EEG. This could be addressed by adding more hypothesis RDMs reflecting different aspects of the images that may correlate with real-world size.

- The choice of ANNs lacks a clear motivation. Why these two particular networks? Why pick only 2 somewhat arbitrary layers? If the goal is to identify more semantic representations using CLIP, the comparison between CLIP and vision-only ResNet should be done with models trained on the same training datasets (to exclude the effect of training dataset size & quality; cf Wang et al., 2023). This is necessary to substantiate the claims on page 19 which attributed the differences between models in terms of their EEG correlations to one of them being a 'visual model' vs. 'visual-semantic model'.

- The first part of the claim on page 22 based on Figure 4 'The above results reveal that real-world size emerges with later peak neural latencies and in the later layers of ANNs, regardless of image background information' is not valid since no EEG results for images without backgrounds are shown (only ANNs).

Appraisal of claims:

While the method shows useful and interesting patterns of results can be obtained by combining contrasting behavioral/image metrics, the lack of additional control models makes the evidence for the claimed unconfounded representation of real-world size in EEG responses incomplete.

Discussion of likely impact:

The paper is likely to impact the field by showcasing how using partial correlations in RSA is useful, rather than providing conclusive evidence regarding neural representations of objects and their sizes.

Additional context important to consider when interpreting this work:

- Page 20, the authors point out similarities of peak correlations between models ('Interestingly, the peaks of significant time windows for the EEG × HYP RSA also correspond with the peaks of the EEG × ANN RSA timecourse (Figure 3D,F)'. Although not explicitly stated, this seems to imply that they infer from this that the ANN-EEG correlation might be driven by their representation of the hypothesized feature spaces. However this does not follow: in EEG-image metric model comparisons it is very typical to see multiple peaks, for any type of model, this simply reflects specific time points in EEG at which visual inputs (images) yield distinctive EEG amplitudes (perhaps due to stereotypical waves of neural processing?), but one cannot infer the information being processed is the same. To investigate this, one could for example conduct variance partitioning or commonality analysis to see if there is variance at these specific time-points that is shared by a specific combination of the hypothesis and ANN feature spaces.

- Page 22 mentions 'The significant time-window (90-300ms) of similarity between Word2Vec RDM and EEG RDMs (Figure 5B) contained the significant time-window of EEG x real-world size representational similarity (Figure 3B)'. This is not particularly meaningful given that the Word2Vec correlation is significant for the entire EEG epoch (from the time-point of the signal 'arriving' in visual cortex around ~90 ms) and is thus much less temporally specific than the real-world size EEG correlation. Again a stronger test of whether Word2Vec indeed captures neural representations of real-world size could be to identify EEG time-points at which there are unique Word2Vec correlations that are not explained by either ResNet or CLIP, and see if those time-points share variance with the real-world size hypothesized RDM.

Reviewer #3 (Public Review):

The authors used an open EEG dataset of observers viewing real-world objects. Each object had a real-world size value (from human rankings), a retinal size value (measured from each image), and a scene depth value (inferred from the above). The authors combined the EEG and object measurements with extant, pre-trained models (a deep convolutional neural network, a multimodal ANN, and Word2vec) to assess the time course of processing object size (retinal and real-world) and depth. They found that depth was processed first, followed by retinal size, and then real-world size. The depth time course roughly corresponded to the visual ANNs, while the real-world size time course roughly corresponded to the more semantic models.

The time course result for the three object attributes is very clear and a novel contribution to the literature. However, the motivations for the ANNs could be better developed, the manuscript could better link to existing theories and literature, and the ANN analysis could be modernized. I have some suggestions for improving specific methods.

(1) Manuscript motivations
The authors motivate the paper in several places by asking " whether biological and artificial systems represent object real-world size". This seems odd for a couple of reasons. Firstly, the brain must represent real-world size somehow, given that we can reason about this question. Second, given the large behavioral and fMRI literature on the topic, combined with the growing ANN literature, this seems like a foregone conclusion and undermines the novelty of this contribution.

While the introduction further promises to "also investigate possible mechanisms of object real-world size representations.", I was left wishing for more in this department. The authors report correlations between neural activity and object attributes, as well as between neural activity and ANNs. It would be nice to link the results to theories of object processing (e.g., a feedforward sweep, such as DiCarlo and colleagues have suggested, versus a reverse hierarchy, such as suggested by Hochstein, among others). What is semantic about real-world size, and where might this information come from? (Although you may have to expand beyond the posterior electrodes to do this analysis).

Finally, several places in the manuscript tout the "novel computational approach". This seems odd because the computational framework and pipeline have been the most common approach in cognitive computational neuroscience in the past 5-10 years.

(2) Suggestion: modernize the approach
I was surprised that the computational models used in this manuscript were all 8-10 years old. Specifically, because there are now deep nets that more explicitly model the human brain (e.g., Cornet) as well as more sophisticated models of semantics (e.g., LLMs), I was left hoping that the authors had used more state-of-the-art models in the work. Moreover, the use of a single dCNN, a single multi-modal model, and a single word embedding model makes it difficult to generalize about visual, multimodal, and semantic features in general.

(3) Methodological considerations
a) Validity of the real-world size measurement
I was concerned about a few aspects of the real-world size rankings. First, I am trying to understand why the scale goes from 100-519. This seems very arbitrary; please clarify. Second, are we to assume that this scale is linear? Is this appropriate when real-world object size is best expressed on a log scale? Third, the authors provide "sand" as an example of the smallest real-world object. This is tricky because sand is more "stuff" than "thing", so I imagine it leaves observers wondering whether the experimenter intends a grain of sand or a sandy scene region. What is the variability in real-world size ratings? Might the variability also provide additional insights in this experiment?
b) This work has no noise ceiling to establish how strong the model fits are, relative to the intrinsic noise of the data. I strongly suggest that these are included.