- Rating
- 3(+1)
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
In this paper the authors introduce SEIS, a novel subspace-based metric for estimating invariance and equivariance of neural representations without external labels. The metric separates the neural activations into "spatial" and "observational" axes and assembles them in matrix-shaped form. An SVD decomposition is then used to provide directions weighting and compute both invariance and equivariant agreement scores. The author then compute the final metric by a final chance-correction step that accounts for the expected random representation overlap. The authors then first validate the metric on a synthetic benchmark of known transformations and later deploy it to study the neural representations in different architecture and learning regimes: they show that both invariance and equivariance build-up during training, that augmentation mainly operates on deeper layers and that both loss and architecture decision influence the final score. Finally, they measure the correlation between the label-free estimation of representation equivariance and model performance on affine-transformed inputs.
Critical strengths
The authors introduce a novel metric to estimate invariance and equivariance of neural representations which is easy to understand and efficient to compute, while requiring no external labels and across the networks' depth. This could be a valuable addition to the toolkit available to study network representations, especially as they evolve during training and architecture choices. The authors then offer several use-cases for the application of the novel metric to the study of representation in various networks and tasks to illustrate this point. Lastly, the paper is generally very well written and easy to follow, with clear illustrations.
Critical weaknesses
One important limitation of the current work is a lack of comparisons between the authors' proposed metric and the alternatives beyond the synthetic benchmark of Fig.1. The authors should report, when feasible, the equivalent scores obtained with the alternative technique to assess whether they agree of not, and in case discuss these differences and the relative advantages. A second important aspect that is missing is a more careful characterisation of the metric as a function of the transformation severity: does the metric report even for rotations of ? Or does it offer more gradual responses? The current synthetic benchmark conflates this distinction and only reports the overall score distribution across transformation type.
Questions for authors
Below are some relevant questions for the authors. I would be willing to increase my score past acceptance level even if only part of the following list is adequately addressed.
- In Fig. 1a, what is the expected ground-truth score/distribution?
- In Fig.1a authors tests several transformations, however it remains unclear whether their metric is capable of tracking the severity of the transformation, as the chosen visualisation only shows the overall spread. However it would have been interesting to see whether the metric was capable of assigning calibrated graded scores for invariance separating near-identity transformations (e.g. scaling ~0.95, rotation ~5°) from more severe changes. Having a smaller variability than Lenc-Vedaldi does not automatically mean a better metric in this regard. (Fig. 8 partially reports the scores for different rotations [90°, 180°, 270°], but critically misses the ~identity range)
- In Fig.1b, can the author report the convergence curves for the other transformations?
- Related to Fig. 2, can the author report the per-layer curves for both and ? This would directly support the authors claim that "[...] deep representations converge progressively from lower to higher layers during training".
- Why did the author chose to report instead of directly ? This would have allowed for example in Fig.9 to directly compare whether the from-scratch models achieved comparable level of invariance/equivariance relative to their pre-trained counterparts.
- In Fig. 5 (but also elsewhere), can the author report the error-bars (std/sem) over the explored 5 seeds in their visualisations to assess whether the observed difference between positional encoding schemes is significant.
- In Fig. 5, can the authors discuss why they think the NoPE configuration seems to perform worse despite the fact that no positional information is provided, thus forcing the transformer to use local image context similarly to CoPE?
- In Fig. 6b,d there seems to be a visual split between two groups of "good performer" (high or with relatively lower ) and "bad performer" (the opposite), which lowers the overall correlation between the validation score and the model performance. Do the authors observed some model characteristics that might explain this fact? (And can the authors also report the correlation between and ?)
- Can the authors report what the alternative metrics (L-V, G-EED) measure for their use-case experiments (e.g Fig. 3, Fig.4, Fig.5)? Do the metrics agree in the reported invariances?
- In Fig.3, the fact that both invariance and equivariance decrease in the initial layers and then increase in deeper ones is reminiscent to the fact identified in previous works (Ansuini et al. NIPS 2019; Muratore et al. NIPS 2022; Valeriani et al. NIPS 2023) that the intrinsic dimension of the data manifold first increases and then decreases, which was directly related to the initial suppression of "unwanted" regularities (e.g. luminance/contrast) followed by a subsequent buildup of more abstract and task-relevant invariances (e.g. cat/dog). It would be perhaps interesting for the authors to discuss this potential connection in their work.
4 comments
Q9. Lenc–Vedaldi and G-EED on real-network use cases
Thank you for your suggestion. We applied Lenc–Vedaldi and G-EED to the same models and transformations for comparison.
For Fig. 3, both baselines confirm that augmentation strengthens deep-layer equivariance and invariance.
For Fig. 4 & Tab. 1: Within each architecture, both baseline rankings correlate positively with SEIS. Across architectures, SEIS favours UNet for both scores, Lenc–Vedaldi favours FCN, while G-EED favours UNet for equivariance but FCN for invariance. We additionally correlated every score with mIoU as shown in table below. SEIS is the only method for which both equivariance and invariance are positively associated with mIoU. Its advantage here is the consistent behaviour of both disentangled scores across architectures. This connection is consistent with prior segmentation work showing that equivariance/invariance can provide useful pixel-level inductive biases.
| Methods | ||
|---|---|---|
| 0.92 | 0.88 | |
| 0.93 | 0.87 | |
| -0.81 | -0.59 | |
| -0.58 | -0.34 | |
| -0.83 | -0.66 | |
| 0.94 | 0.87 |
For Fig. 5, Lenc–Vedaldi reproduces exactly the SEIS ordering for both scores. G-EED also strongly agrees.
For Fig. 6, we evaluated Lenc–Vedaldi and G-EED using the same 50 models, same ImageNet validation images, final spatial representations, and fixed affine transformation as Sec. 4.5.
| Methods | - | |
|---|---|---|
| 0.90/0.92 | 0.62/0.77 | |
| 0.74/0.80 | 0.77/0.82 | |
| 0.60/0.66 | 0.61/0.70 | |
| 0.14/0.23 | 0.05/0.17 |
Each cell reports Pearson ()/Spearman (). best predicts absolute transformed accuracy, while SEIS is the strongest representation-based indicator for and the strongest overall indicator for accuracy retention. We therefore interpret SEIS as providing complementary label-free robustness information.
We will add these comparisons in the revised paper with clear visualisations.
Q10. Connection to intrinsic dimension
Thank you for this interesting connection. However, we believe they measure different properties. Intrinsic dimension quantifies manifold complexity, while SEIS measures alignment between the paired manifolds. Therefore, low intrinsic dimension does not imply high equivariance or invariance, or vice versa. For example, a high-dimensional feature map can be perfectly equivariant if it simply shifts with the input.
I thank the authors for the detailed rebuttal and the new experiments that have addressed most of my concerns (e.g. per-layer dynamics, baseline comparisons etc...), although some of the mentioned results like the Lenc-Vedaldi/G-EED comparisons mentioned in the rebuttal (Fig.3 and 5) are not in the revised paper (or I couldn't locate them). I have raised my score to 3 to reflect the improved manuscript.
Thank you for taking the time to review our rebuttal and revised manuscript. The comparisons have now been incorporated into the latest revision. The augmentation comparisons appear in Fig. 3, and the positional-encoding comparisons corresponding to Fig. 5 appear in Appendix H.
We thank the reviewer for their thorough review and valuable feedback. Below, we address each question point by point.
Q1. Ground-truth interpretation for Fig. 1a
The synthetic construction provides the following expectations: (1) Identity: both scores should be 1; (2) Small noise: both scores should remain near 1; (3) Spatial transformations: Sequiv=1, there is no universal scalar ground truth for Sinv, near-identity transformations should score close to 1, while stronger transformations should produce lower Sinv ; (4) Independent activations: both scores should be 0.
Q2. Transformation severity for Fig. 1a
Thank you for pointing it out. For the new experiment, we varied rotation from 0 to 180, scaling by up to 20%, and translation by up to 15% of image width, using 50 trials per severity. SEIS invariance decreases consistently with transformation severity, with Spearman correlations of ρ=-0.98 for rotation and ρ=-1 for both translation and scaling. These results confirm that SEIS provides a graded response, including near identity.
Lenc–Vedaldi also exhibits a graded response but requires learning a transformation-specific latent map, whereas G-EED is less sensitive to increasing severity and requires prior knowledge of the transformation action to construct the expected equivariant target. SEIS requires neither.
We will add these results into the revised paper.
Q3. Convergence across transformations for Fig. 1b
We will add separate convergence curves for translation, scaling, and rotation. The same convergence pattern holds for all three transformations. Our practical recommendations remain unchanged.
Q4. Per-layer training curves for Fig. 2
Thank you for this suggestion. We initially omitted the per-layer curves because plotting every layer would introduce too many overlapping lines and reduce readability. In the revised paper, we will group layers by network blocks to provide a clear view of how the representations evolve from shallower to deeper blocks during training.
Q5. Why report ∣ΔS∣ instead of S in Fig. 2
We report ∣ΔS∣ because this section asks how much training changes each representation relative to its own initialisation. This is appropriate when comparing pertained and scratch models, whose initial scores may differ substantially. Comparing raw S would mix the model’s initial scores and the change caused by training.
Q6. Error bars over seeds in Fig. 5
Thank you for this suggestion. We will redraw Fig. 5 with mean and SD across five seeds. The resulting error bars are generally small and preserve the qualitative ordering across transformer blocks. Conditional PE yields the highest Sequiv and Sinv, Relative PE is intermediate, and Absolute/No PE are lower.
Q7. Why can NoPE perform worse than CPE in Fig. 5?
The premise that removing positional encoding forces the transformer to use local context is not correct. Without positional encoding, self-attention remains global, so it cannot distinguish neighbouring from distant patches. CPE instead uses a shared depth-wise convolution to encode local spatial information while preserving translation structure.
Q8. Two model groups in Fig. 6 and Sequiv–Aval correlation
Thank you for this observation. We analysed the split separately in both Fig. 6b and Fig. 6d. In both figures, a two-component Gaussian mixture is preferred to a single component. However, we do not find that the split can be explained by architecture type. Instead, we found out that models with similar validation accuracy can differ greatly in transformation equivariance. The higher- and lower-robustness components have mean Sequiv of 0.78 and 0.57, respectively.
The Spearman between Sequiv and Aval is 0.63, this moderate relationship confirms that the indicators share some signal but are not redundant. SEIS is not merely reproducing validation accuracy, it can be regarded as a complementary information when both indicators are available, and enables label-free robustness when validation labels are unavailable.