acceptodds
Under review as a conference paper at ICLR 2027

Age Subspaces in 3D Face Representations: Readability, Predictive Dependence, and Transferability

Abstract

Understanding age-related face representations requires distinguishing whether age information is linearly readable, whether a fitted predictor relies on the identified directions, and whether those directions transfer across independently trained representations. We investigate these questions in 3DMM-derived shape representations using correlation-ranked PCA directions, controlled direction removal, and cross-run comparisons. Correlation-ranked PCA directions retain age-predictive performance with fewer dimensions than variance-ranked PCA directions. Under the evaluated removal protocol, removing the top ten correlation-ranked directions reduces the fixed predictor's from 0.588 to approximately 0.16, whereas dimension-matched random removal reduces by only 0.017. Across training runs, the discovered subspaces exhibit overlap close to the random baseline in native coordinates, and directly transferring these directions across runs produces limited removal effects. These findings establish predictive dependence within the measured representations without implying complete age-information removal or intrinsic semantic differences across models. We further introduce Age Subspace Learning (ASL), a PCA-based decomposition with identity supervision, for age-robust face representation. Together, the results distinguish age readability, predictor dependence, and direct transferability, clarifying the interpretation and scope of age-subspace analysis.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Representation Learning, Age Subspace Analysis, 3D Face Representations, Linear Probing, Predictive Dependence, Cross-Representation Transferability

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_nCmQOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

The paper analyzes the latent space of a 3D face model (encoder). More specifically, it decomposes the space using PCA and analyzes the influence of each component on the readability, predictability, and transferability of facial age.

Critical strengths
  • Relevance of the Problem: The study of how semantic attributes are encoded and geometrically organized in latent spaces is a topic of great interest to the research community.
Critical weaknesses
  1. Failure to distinguish between removing age-specific information and destroying the overall signal

    Throughout the manuscript, it is claimed that the drop in the fixed predictor’s performance when directions are removed demonstrates "predictive dependence beyond correlation" and the "functional importance" of those directions. However, the experimental protocol does not demonstrate this. By removing principal or highly correlated dimensions without retraining or evaluating what happens to the rest of the facial information, the entire signal of the representation, not just age, is completely degraded. If the overall signal is destroyed, all the criteria proposed by the authors (a drop in R2R^2 relative to a random control) are trivially met without this implying that a semantic subspace of age has been isolated.

  2. Limitations and disclaimers that render contributions meaningless

    In the Discussion and Conclusion sections, the article repeatedly emphasizes what its experiments do not demonstrate: that the drop in R2R^2 does not imply that age information has been removed from the residual representation, that the low transfer between runs does not imply that the semantics are different because the coordinates need to be aligned, and that reducing the age prediction does not imply improved identity preservation nor does it demonstrate semantic disentanglement. If the experimental protocol leaves all these fundamental questions open and the observed effects are direct consequences of degrading the signal or of failing to align independent coordinate systems, what does the presented empirical analysis actually contribute or demonstrate?

  3. Selection of PCA for the candidate directions

    The analysis uses PCA to generate an orthonormal set of candidate directions, which are then ranked based on their absolute correlation with age. Why is PCA specifically chosen to generate the initial candidate directions if the subsequent goal is to search for an association with age?

MINOR

  1. Origin of the 3DMM representation

    When formulating the problem, it starts directly with a representation derived from 3DMM with 80 dimensions z∈R80z \in \mathbb{R}^{80}. Where exactly does this representation come from?

  2. Model trained and loss functions

    When introducing the function loss Ltotal=Ltask+λidLid+λageLageL_{total} = L_{task} + \lambda_{id}L_{id} + \lambda_ {age}L_{age}, where Ltask=Lreg+Lrec+LpercL_{task} = L_{reg} + L_{rec} + L_{perc}, it is not clear which specific model or modules are being trained using this loss function, nor how each of these terms is mathematically defined.

  3. Results charts

    Figure 4 shows three panels (R2R^2, MAE, and “Causal Effect”). The three panels essentially show the same redundant information (especially the first and third, where ΔR2\Delta R^2 is simply the same curve as in the first panel shifted by a constant; furthermore, the term “Causal Effect” is used in the figure’s title even though the text explicitly states that claims of causality should be avoided).

  4. Comparison with the state of the art

    Table 2 ("Comparison with state-of-the-art methods") lists numbers alongside methods such as ArcFace, MagFace, and AdaFace without specifying in the title or in the table headers which metric is being reported or what the exact task being evaluated is. Is this the state-of-the-art for facial recognition measured in verification accuracy (%)?

Questions for authors

The paper’s main methodological problem is that it does not demonstrate that age-related information has been isolated, but rather that the signal is being degraded. To demonstrate that the loss of information is age-specific and not a general degradation of the representation, it is necessary to simultaneously measure the extent to which performance declines in other attributes.

Reviewer_BKZ1OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work aims to distinguish three complementary properties of the age subspace in 3D face representation: readability, predictive dependence, and transferability. This work also found that correlation-ranked PCA directions capture age information efficiently and are important for prediction but show limited transferability across independently trained models. A PCA-based decomposition, age subspace learning (ASL), is introduced for age-robust face representation.

Critical strengths
  1. Three complementary research questions are carefully raised and studied; however, I am not certain that these three concepts: readability, predictive dependence, and transferability, have not been uniformly studied before. This would be a significant advantage of this work if it hadn't existed before.
  2. This work provides a detailed mechanism analysis and ablation study. Table 2 also shows the improvement over previous baselines on five mainstream datasets. However, I find it difficult to ensure the reliability of these results because, although the original paper mentions that the code will be released, as stated in the reproducibility statement, some auxiliary components may not be included in the public release.
Critical weaknesses
  1. The caption of Figure 1 in the paper claims that the generated faces show progressive age-related appearance changes, but for the faces in the second row, the 45-year-old appears younger than 40, with fewer wrinkles and darker hair. This seems contradictory.
  2. The last sentence of the fifth paragraph of the introduction: "without assuming that..."; the last sentence of the caption in Figure 1: "it is not evidence that..."; the last sentence of the first paragraph of Section 3.3: "we do not interpret..."; the last sentence of the caption in Figure 2: "it does not assume...", etc. These defensive writing elements appearing in the main text could be strange and distracting for the reader, perhaps a result of the LLM revision.
Questions for authors

Clarification Requests:

  1. If the code for this work can be made as open source as possible, readers would have more confidence in the original content.
  2. These two critical weaknesses mentioned above might make readers feel that the paper is not rigorous and raise doubts about the research content, even if this is not the case.
Reviewer_rqyyOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies how chronological age is encoded in an 80-dimensional 3DMM-derived shape representation. It separates three questions: whether age can be read out linearly, whether a fixed age predictor depends on the identified directions, and whether those directions transfer across training runs.

PCA directions ranked by their correlation with age reach 80% and 90% of the full R^2 with 13 and 23 dimensions. Variance ranking needs 20 and 37. Removing the top ten correlation-ranked directions lowers the fixed predictor's R^2 from 0.588 to about 0.16. Removing random directions of the same dimension changes it by only 0.017. Across seeds, the subspaces overlap at chance level in native coordinates, and a subspace transferred from another run has almost no effect.

The paper also proposes Age Subspace Learning (ASL), which projects out the first ten variance-ranked PCs and maps the remainder to an ArcFace embedding. ASL is reported to outperform ArcFace, MagFace and AdaFace on five face verification benchmarks.

Critical strengths

Separating readability, dependence of a fixed predictor, and recoverability by a retrained predictor is a useful way to frame this kind of analysis. The removal experiment also has the right controls: random subspaces of matched dimension with error bars, and variance-ranked removal. The paper is honest about its limits, for example that it does not test transfer after Procrustes alignment. Together these give a reasonable basis for a good paper.

Critical weaknesses

The main concern is that the core diagnostic results are close to true by construction, and the protocol is not described in enough detail to rule out selection effects. Ranking features by their correlation with the target and then measuring how fast R^2 accumulates is greedy supervised feature selection, so it will beat variance ordering almost by definition. Likewise, removing the inputs most correlated with age from a fixed regressor will hurt it more than removing random inputs.

The paper does not state whether ranking, fitting and evaluation use separate samples. With "CelebA-200" and "AgeDB-400" and an 80-dimensional input, in-sample R^2 would be inflated, and same-sample ranking would favor the correlation ordering further. Several other details are also unclear:

  • The source of the chronological age labels. CelebA appears to provide only a binary "Young" attribute, and the paper cited for CelebA-200 is a bias-analysis paper.
  • The full-representation R^2, which appears as 0.588, 0.637 and 0.549 in different places without explanation.
  • The predictor, which is called a linear probe in the text but an MLP regressor in Fig. 2.

For RQ3, the interpretation depends on what the 80 coordinates are. If they are coefficients of a fixed 3DMM basis, as the text on p. 5 suggests, the coordinates are already shared across runs. The misalignment caveat would then not apply, and chance-level overlap would itself need an explanation. If they are learned latents, chance-level overlap without alignment is expected, and RQ3 is less informative.

A second concern is Table 2. ASL's identity representation comes from a small network that maps the shape coefficients onto ArcFace features, yet ASL is reported above ArcFace itself on all five benchmarks (99.45 vs 99.20 on LFW). A student that predicts its teacher's embedding from geometry alone and outperforms the teacher needs an explanation. The baseline numbers are also well below the roughly 99.8% on LFW usually reported for these methods. The paper does not give the test-time embedding, the MORPH and CACD protocols, the source of the baseline numbers, or what "3DMM-Baseline" is.

Sec. 4.4 refers to ablations over k, PC1-only removal and identity supervision in Appendix A.3, but A.3 only describes them and it doesn't refer to any Tables or Figures in the Appendix. The one numeric ablation, Table 4, has its own gaps. It does not say which benchmark "Accuracy" refers to, and its last two rows differ in a way that is not stated. L_age is also described differently in Sec. 3.3 and in Fig. 2.

Finally, ASL uses the first ten variance-ranked PCs, and several of these are barely correlated with age in Fig. 3. The analysis therefore does not directly motivate the application design.

Questions for authors
  1. How many samples are used in each analysis, and where do the age labels come from? Are PCA, ranking, predictor fitting and R^2 evaluation done on disjoint splits? What are the held-out R^2 values?
  2. Is the age predictor linear or an MLP? Which setting produced 0.588, 0.637 and 0.549?
  3. Are the 80 dimensions coefficients of a fixed 3DMM basis, and if so which one, or are they learned latents? What differs between seeds?
  4. For Table 2:
    • Which embedding is used at test time?
    • What protocol is used for MORPH and CACD?
    • Where do the baseline numbers come from?
    • How does h(z_comp) end up above its ArcFace teacher?
  5. Where are the ablation results referred to in Sec. 4.4 and Appendix A.3? What separates the last two rows of Table 4?
  6. What generated the photorealistic images in Fig. 1 from a shape-space direction? In the top row, identity appears to change noticeably between 40y and 45y. In the second row, face appears to look younger from 40y -> 45y.
  7. A retrained probe on the residual representation, as described in Sec. 3.4, would show whether the R^2 drop reflects information that was actually removed or a brittle fixed predictor.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.