- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper proposes measuring model–brain alignment in a shared latent space learned by a multimodal VAE. The authors evaluate distributional overlap and image identification in this space, using ResNet representations and fMRI data. They test the approach on constructed representations that remove image details, swap category information, or rely on artificial shortcuts. The proposed scores generally penalize these constructions more strongly than the existing measures considered.
Critical strengths
I like the use of constructed controls. The networks trained to classify using planted squares provide a useful test of whether an alignment measure rewards a model that solves the task through an implausible strategy. The evaluation on held-out images and checks across training seeds, feature selections, and RSA variants also give some confidence that the reported patterns are reproducible.
Critical weaknesses
-
I am not fully convinced that the corruption ladders establish a correct ordering that an alignment measure should recover. Removing image-specific detail loses information, but a fixed, one-to-one category swap can preserve category identity in a different code. A fitted decoder may reasonably undo that mapping. The constructions combine these different changes, and I think the authors need to explain more clearly which properties their measure should preserve or penalize. Positive controls using equivalent representations would also help establish that the method does not reject harmless changes.
-
I am unsure whether the results establish an advantage over existing approaches. Appendix F reports that ridge already ranks every foil below both reference networks and detects every step of the corruption ladders. At each network’s best layer, RSA and CKA also rank every foil below the reference. Larger numerical differences on a different score scale do not by themselves establish a better measure. Meanwhile, a separate latent space is trained for each candidate model, so the brain representation used for evaluation changes across comparisons. I would like to understand how stable the conclusions are to latent dimension and reconstruction-loss weighting, and how much the VAE adds over simpler bidirectional or shared-space methods.
Questions for authors
-
What transformations should leave the proposed alignment scores unchanged?
-
Does the advantage remain against reverse predictivity or a simpler shared space method evaluated using the same image identification task? How sensitive are the rankings to latent dimension and the balance between model and brain reconstruction losses?
-
The dCKA estimator corrects stimulus sampling but not feature sampling. Could the authors discuss whether the correction in Chun et al. (2025, arXiv:2502.15104) is applicable to their fMRI measurements and projected model features? I am unsure whether it would affect the comparisons, especially if only the brain side requires correction.
4 comments
Response to Reviewer ZxfB (2/4)
2. i) Advantage over existing approaches
We do not claim that our method is superior because of larger differences on another scale, but that it expresses different aspects of alignment, i.e., distributional and stimulus-specific, that the standard metrics do not capture. This leads to the foil models scoring low in our method's metrics, while they reach most or all of the reference RN50's score on the standard measures.
Specifically, the issue that we show with the paper’s constructed cases and RSA’s and ridge’s performance on them is that the foils still manage to get a high score. Even in Ridge, the foils score up to 77% of RN50’s and 82% of RN50sc’s r, and 61% of RN50’s, and 71% of RN50sc’s and they are scored at the level of and . For the per-layer RSA, both and score above both references, and Carrier-B is level with RN50sc (88% and 87% of RN50 on RSA; Appendix N, Table 19). The table below places the best foil score on each measure as a % of what score RN50 achieves in the same metric.
| measure | best foil, % of RN50's score | where it falls on the training-degradation ladder |
|---|---|---|
| RSA | 121% | above the trained RN50 |
| ridge r | 77% | above α = .25 |
| ours, top-1 | 9% | at the untrained network |
If foil model representations that haven’t been fed the images, or solve the task in mechanistically dissimilar ways with the brain (e.g., from planted pixels shortcuts) can achieve most of RN50’s and RN50sc’s alignment score and all of and (RN50 representations that have some information removed, but still contain substantial RN50 information), doesn’t this make RN50 and RN50sc high alignment score less meaningful?
In the distributional and stimulus-specific alignment that our method measures, RN50’s and RN50sc’s score cannot be achieved with a foil. So, our thesis is not that our method is superior, but that it expresses different aspects of alignment, which in some cases are more difficult to achieve a high score on, and this is why the foil models fail. We believe these aspects are meaningful and could potentially be what is needed to distinguish candidate models (architectures, training objectives etc.) regarding their alignment with the brain in cases where other metrics saturate.
2. ii) Comparison with CCA (added in Appendix I)
CCA also places model and brain in a common space and in addition to its own metric (canonical correlation r), we can compute the metrics we proposed in this work in their aligned space. CCA common space identifies images well (RN50 top-1 0.568; full results in Table 9, latent visualisation of CCA space in Fig. 8, App. I). However, CCA whitens each side (model and brain). Every retained direction (after PCA cut) is rescaled to unit variance, so how much of the model representation a direction carries is discarded, and only the directions that correlate best with the brain are kept in the common space. Thus, CCA gives a model credit for whichever part of it correlates best with the brain, even when that part is only a small fraction of the model's representation. Our method works differently, since the latent space must also reconstruct the model's features (model→model and brain→model pathways), so a part of the model that carries a lot of its variance needs to be represented in the common space.
To test exactly that, namely if CCA score can be maximised if a model representation contains a brain-like subspace, no matter how small the part of the model representation that aligns with the brain is, in Appendix I.2, we constructed representations in which the model representation is controlled to be consisted of a meaningful RN50 signal and Null’s signal. We take Null (RN50's response to a different image) and add RN50's response to the true image at a gain that sets its share of the variance. When the true image's RN50 code is 8% of the mixed representation and the other 92% is about another image (Null), CCA scores it at 56% of the top-1 and at 82% on its own measure r of what RN50 itself achieves; when RN50 code is 11% of the share, at top-1 is at 80% and r at 93% of what the actual RN50 representation achieves. Our method's top-1 metric scores both at 1% of RN50, and rises only when the code is a large part of the representation (top-1 at 30% of RN50 at a 38% share, and RN50's level at 71% share) (Table 11, App. I.2). The reason is whitening: CCA rescales every retained direction to unit variance, so a direction that holds 1% of the variance counts as much as one that holds 50%. This is a known property: a measure that is invariant to invertible linear maps ignores the scale of directions and gives the same score to representations that match only in their small principal components (Kornblith et al., 2019). Our latent must also reconstruct the model's features, so a code enters it in proportion to its share.
Response to Reviewer ZxfB (3/4)
2. iii) Impact of model vs. brain loss weights and Latent dimensions number
In our method, every training image pairs the model features with the one subject who saw it, so the model encoder sees all 62,551 training images and each subject's encoder 7,400–8,100. Thus, each subject's own encoder and decoder receives fewer examples than the shared model encoder/decoder. However, over the eight subjects the brain contributes as many examples as the model, since each image has been seen by exactly one subject. Thus, at least in respect to the number of the samples, the model and brain representation samples equally contribute to shaping the latent space.
Yet, there are other imbalances on the contribution of model vs. brain in shaping the latent space that come from training settings. In the per image loss, the model features carry ~51k variance units (model features are z-scored with variance 1) and the voxels ~11k on average (after the NSD and Algonauts preprocessing the brain features have variance ~0.5), so the model loss terms outweigh the brain and the model is effectively used as a scaffold in the latent space as the Reviewer suggests.
In Appendix H, to explore the effect of the model vs. brain losses in shaping the latent space and the sensitivity of our method’s metrics to the weight of the model vs. brain reconstruction losses, we retrained VAE with the two sides balanced (weights 1 for the model terms and 4.52 for the brain terms, KL weight scaled to keep its share comparable to the original one), everything else as in the paper. With this weighting the model and brain reconstruction losses contribute equally in variance. Table 8 in Appendix H gives the metrics for the paper’s main analysis vs. balanced weights comparison side by side and the visualization of the latent is in Fig. 7.
With balanced weights the brain shapes the latent space more than before, and the level of the scores changes. Identification of the references falls by almost half (RN50 top-1 0.321 to 0.179), because the latent now spends capacity on brain signal that the model does not explain. However, the ranking between the different cases of the main analysis (RN50, RN50sc, Ladders, and Foil models) remains unchanged (Table 8, App. H). On top-1, top-5, median rank and iLISI the order RN50 > RN50sc > , > foils is preserved. Despite the margins being narrower than with the paper's main analysis weights, the ordering does not depend on the model being the anchor in the latent space.
The ordering is the same with both weightings, so our conclusions do not depend on this choice. We view the weights of the model and brain losses as a hyperparameter, similarly to the latent dimensionality: when comparing different models regarding their alignment to a brain dataset, they should be compared with the same loss balancing between model and brain representations. We also view this parameter as a design choice, because it has a clear meaning: it sets how much the brain, relative to the model, shapes the common space. With the paper's weights the model works as a scaffold. With balanced weights, more brain signal is hold in the latent.
We note that the scaffold cannot create alignment by itself. With the correspondence between images broken (Null), every latent read-out of our method is at chance.
Latent dimension (Appendix G, Tables 6–7). After the reviewer's suggestion, we trained our VAE with different latent dimensionalities for RN50, RN50_sc, , , the three highest performing foils —Carrier-B, Carrier-P and LabelCode-N—and the Null. The ranking of these remained the same in 64 and 128 dimensions as it was in the originally trained 32-d VAE.
With 64 and 128 dimensions instead of 32, no representation moves by more than 0.02 on top-1, the order remains the same and the ordering contrasts remain significant on identification. The energy distance is built from Euclidean distances between latent points, and these grow with the number of dimensions over which they are summed. iLISI counts how well model and brain points are mixed among each point's nearest neighbours; with the same 872 images spread over more dimensions the neighbourhoods become sparser, and a small offset between the two clouds in each dimension adds up, so the neighbours of a point are more often from its own cloud. The values of these two metrics should therefore, because of nature of the metric, be compared between different model representations at one latent size, and not between latent sizes. Within each latent size the order RN50 > , > foils is preserved on both metrics (energy and iLISI), and all eight contrasts are significant at 32, 64 and 128 dimensions. The one change is on iLISI at 128 dimensions, where RN50sc (0.360) falls just below (0.379).
Response to Reviewer ZxfB (4/4)
Question 1. What transformations should leave the scores unchanged? (analysis added in Appendix K).
Changes that only re-express a representation should leave the scores unchanged (Appendix K, Table 15). We tested the following:
- The training seed of our VAE: Over three seeds, top-1 of a representation varies by at most 0.007 (0.012 for RN50). This range is the reference for the two tests below.
- A rotation of the feature space that keeps every feature's variance (we tested it in the following representations: RN50, and Carrier-P): For the rotation of the feature space, identification stays inside the range of the three VAE training seeds or within 0.010 of it (RN50 top-1 0.329 against 0.321–0.333), and the order RN50 > > Carrier-P is kept on every read-out in all eight subjects. iLISI is inside the range for RN50 and Carrier-P and above it for (0.536 against 0.460–0.496).
- An independently trained copy of a network (RN50sc retrained from another initialisation): Every read-out is inside the range of training seeds (top-1 0.193 against 0.192–0.199).
What should lower the scores, and does: removing or replacing information about the image (the two Corruption ladders), undoing the network's training (the training-degradation ladder), and breaking the correspondence between images (the Null, at chance).
Lowered, as they should be: removing or replacing information about the image (the two Corruption ladders), undoing the network's training (the training-degradation ladder), and breaking the correspondence between images (the Null, at chance).
Question 2. Reverse predictivity, simpler shared-space methods, latent size and loss balance
Reverse predictivity does not keep the ordering. Ridge from voxels to model features rewards representations that are easy to predict: on the real brain the five foils score 2.0–2.7 times RN50 (0.43–0.57 against 0.22), and on the model-recovery study in Appendix Q, where synthetic brain is generated from RN50, the generator RN50 is ranked 11th of 18 (Tables 21–22). The comparison with a simpler shared-space method (CCA) is described above in Section 2.ii (Response 2/4) and in Appendix I. Latent size and loss balance analyses are described above in Section 2.iii (Response 3/4) and in Appendices G and H.
Question 3. Feature-sampling correction of Chun et al. (2025)
We applied feature-sampling correction to either the brain side or both sides, and it changes no comparison. In the estimator of Chun et al. (2025), only each system's self-term is corrected, by a factor that shrinks with the number of sampled features (about 20,700 voxels; 61,548 projected features).
- Applied to brain side alone: dCKA rises by 0.03–0.06% for every representation.
- Applied to both sides: at most 0.7% in the whole-network bases (RN50 0.1960 → 0.1972) and 2.% at the best layer.
- The order of no pair of representations changes.
Response to Reviewer ZxfB (1/4)
We thank the reviewer for the comments. Here are our responses to the two Weaknesses mentioned and the three Questions. Appendix, Table and Figure numbers refer to the revised paper.
1. Corruption ladders and why the measure should penalise them (added Appendix D1).
We agree that a one-to-one swap keeps category identity, and a fitted map can undo it. Here we show that every step down the Corruption ladders removes structure that RN50 shares with the brain: either the image-specific structure within a group, or the similarity structure between groups. The ladders contain two kinds of steps: steps that remove information, and steps that corrupt it, i.e., replace information about an image with the wrong information from another image.
Two different questions can be asked about these manipulations:
i) Can the category of an image still be read out? The answer is yes, as all images of a category still share one mean, and different categories still have different means, so the swap only renames the categories. An invertible linear map can send every swapped mean back to the original one. A measure that fits a linear map therefore cannot see the swap, and ridge does not (0.294 for CATin, 0.293 for CATout, Fig. 3 and Table 5).
ii) Are the same categories still alike? The answer is no. In RN50 some categories are close to each other and others are far apart: the images closest to cats are those of dogs, teddy bears, etc. After the random swap the cat images carry the mean of an unrelated category, and the images closest to them are different, e.g., benches, motorcycles, etc. The categories remain distinguishable, but which categories are neighbours has changed. We call this the similarity structure between categories.
Table 3 in App. D1 (also shown below) measures the second question with two numbers, both computed over pairs of categories (e.g. cat and dog).
Without a swap, the categories that are alike in RN50 tend to be alike in the brain (0.63), e.g., in both the categories closest to cats are dogs, teddy bears and bears. The sibling swap keeps most of RN50’s structure (0.77) and most of its agreement with the brain (0.54). The random swap keeps none of either (-0.02): which categories are alike in the manipulated model representations no longer has anything to do with which are alike in the brain, although the category of every image can still be read out by a linear map.
Table 4 in App. D1 in the paper, splits the agreement with the brain value from the Table above into a part between groups (similarities between group means) and a part within groups (similarities between two images of the same group, group mean removed) and shows that each step of the ladders deteriorates at least one of the two. The agreement between the model's and the brain's similarities within groups falls .167 → .033 → 0 on the category ladder and .153 → .043 → 0 on the supercategory ladder. We argue that this supports that any alignment metric should score CATin, CATout and S-CATdonor, S-CATswap lower than the CATown and S-CATown.
The ordering we defend is: references (RN50, RN50sc) > (CATown, S-CATown) > CATin, CATout and S-CATdonor, S-CATswap. While also (CATown, S-CATown) > foils.