|
Latent Feature Alignment: Discovering Biased and Interpretable Subpopulations in Face Recognition Models |
Soundness: 4: excellent
Presentation: 3: good
Contribution: 4: excellent
Rating: 8: accept, good paper
Confidence: 4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. |
The paper proposes **Latent Feature Alignment (LFA)**, a label-free method for directly discovering semantically coherent and interpretable subpopulations in face-embedding space. LFA iteratively estimates a **latent direction** (a weighted average over embeddings with per-identity reweighting) and grows a group by adding the embedding with maximum projection on that direction until a threshold is met (Algorithms 1–2). Claimed benefits over distance-based clustering include (i) higher **semantic coherence** and (ii) **interpretable directions** that align with attributes (age, ancestry, hair/attire). Experiments on **RFW** and **CelebA**, across **ArcFace, CosFace, ElasticFace, PartialFC**, show improved intra-group attribute consistency versus **k-means** and **nearest-neighbor search**, qualitative/quantitative evidence that traversal along learned directions manipulates attributes when decoded via **Arc2Face**, and that LFA-discovered groups surface **higher FMR/EER** than random groups, revealing biased subpopulations.
* **Label-free bias discovery** that avoids predefined attribute taxonomies; practical for audits where annotations are missing or costly.
* **Identity-balanced direction estimate** (inverse-frequency weights) mitigates dominant-identity pull; method is easy to implement.
* **Consistent semantic coherence gains** across datasets and backbones (Table 1).
* **Interpretability**: arc2face traversals and attribute-probability monotonicity demonstrate that discovered directions track human-meaningful factors.
* **Bias signals**: LFA groups show markedly elevated **FMR** relative to random subsets across models (e.g., up to **×100+** in Table 2 entries), with bootstrapped CIs elsewhere.
* **Initialization dependence & stopping**: quality hinges on pre-clustering/components and a threshold τ; guidance for τ selection and sensitivity analysis is limited.
* **Comparative baselines**: Only k-means/NNS are considered; stronger unsupervised baselines (e.g., subspace methods, spectral clustering, density-based clustering, or latent-direction discovery like Householder/LatentCLR) aren’t directly compared.
* **Statistical reporting**: Table 1 lacks uncertainty; bias metrics include some bootstrap CIs, but broader variance reporting (multi-seed; across init graphs; across τ) is sparse.
* **Causal ambiguity**: Biased evidence comes from impostor-shift analyses within discovered groups; no controlled intervention (e.g., reweighting training data along found directions) to show causal linkage.
* **Compute footprint**: No runtime/complexity profile for large-scale audits (projection loops over millions, graph construction). Practical scaling tips are missing.
1. **Sensitivity/robustness.** How sensitive are results to the cosine-graph threshold (0.5), τ, and to the pre-clustering procedure? Please provide curves for semantic coherence vs. τ and FMR vs. τ.
2. **Baseline breadth.** Can you compare against unsupervised **latent-direction discovery** (e.g., Householder projectors, LatentCLR) and **spectral**/**DBSCAN** clustering under matched group sizes?
3. **Uncertainty.** Add **CIs** for Table 1 and variance across different initial graphs and random seeds; include CIs for Table 2 where possible.
4. **Scaling.** What is the empirical runtime/memory for building the similarity graph and iterating projections on RFW/CelebA? Any heuristics (e.g., ANN indexing, batching) you recommend?
5. **From observation to intervention.** Could you run a small **training intervention**: reweight or exclude the top-bias LFA groups, then evaluate whether FMR gaps shrink?
6. **Direction entanglement.** Traversals sometimes affect attributes asymmetrically (e.g., female→male vs male→female). Can you quantify entanglement (e.g., attribute-probability Jacobians) and study orthogonalization across discovered directions? |
Fully AI-generated |
|
Latent Feature Alignment: Discovering Biased and Interpretable Subpopulations in Face Recognition Models |
Soundness: 3: good
Presentation: 3: good
Contribution: 2: fair
Rating: 6: marginally above the acceptance threshold
Confidence: 3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked. |
The authors propose a method to discover subpopulations in the representation space of face recognition models. Notably, the method does not require labelled data. The method leverages the latent directions to create groups with coherent semantic attributes. Experiments on CelebA and RFW showcase the efficacy of the proposed method, while a connection to algorithmic bias is also uncovered .
- The main idea is simple and clearly presented.
- The experimental results highlight the efficacy of the method on RFW and CelebA.
- The method is thoroughly benchmarked across mutliple SOTA face recognition models.
- The visualization of the learned latent directions with arc2face is neat.
- I appreciate the discussion on the limitations of method.
- The proposed method showcases how to discover sensitive attributes in an unsupervised way. It would be valuable to showcase or discuss if the uncovered directions can be used to debias representations and lead to more fair classification.
- It would be interesting to discuss the potential steerability of the discovered feature clusters based on the initial set of images.
- I am missing some related work on similar clustering-based approaches for generative models, e.g., [1]
[1]. Cluster-guided Image Synthesis with Unconditional Models
Minor:
Some typos:
- Fig.1 "attribtues"
- Ln. 295 "agaisnt"
- Ln 301 "to times"
- Ln 420 "geniune"
I would suggest that the authors address the issues raised in the weaknesses section. In particular, the points raised regarding the potential debiasing of representations and steerability. |
Fully human-written |
|
Latent Feature Alignment: Discovering Biased and Interpretable Subpopulations in Face Recognition Models |
Soundness: 2: fair
Presentation: 3: good
Contribution: 2: fair
Rating: 2: reject
Confidence: 4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. |
This paper introduces Latent Feature Alignment (LFA) an attribute-label-free framework to discover semantically coherent and maybe biased groups within the embedding spaces of face recognition models.
It operates by identifying latent directions derived from groups of embeddings and iteratively aligning samples that share strong projections in these directions. The discovered directions are claimed to correspond to interpretable semantic attributes (e.g., age, ethnicity, etc.), enabling bias auditing without the need for explicit demographic labels.
The authors claim that LFA outperforms k-means and nearest-neighbor clustering in semantic grouping, reveals meaningful and interpretable directions, and discovers groups with measurable bias without explicit annotations.
- Addresses the important gap of bias analysis without labeled attributes
- LFA is simple to implement
- Experiments across four different FR models add some robustness to the empirical observations
- The algorithm closely resembles standard projection-based or spherical clustering methods thus theoretical justification is thin
- Reliance on automatically generated attribute labels from VLMs introduces uncontrolled bias
- Small group sizes and missing statistical tests reduce credibility of claims about discovered disparities.
- Claims about discovering interpretable directions aligned with demographic and contextual attributes extend beyond what is empirically shown
- How does LFA differ fundamentally from spherical k-means or directional clustering beyond the averaging and projection heuristic?
- How stable are the discovered groups under different random seeds or graph connectivity thresholds?
- How reliable are the VLM-generated attribute labels and did the authors manually audit a subset?
- How statistically significant are the FMR differences reported in Table 2 and Table 6?
- Could the discovered latent directions reflect data quality factors (pose, lighting, compression) rather than demographic bias?
- Why does the method stop at a single latent vector per group? Can multiple orthogonal directions be extracted per group? |
Fully AI-generated |
|
Latent Feature Alignment: Discovering Biased and Interpretable Subpopulations in Face Recognition Models |
Soundness: 2: fair
Presentation: 2: fair
Contribution: 2: fair
Rating: 2: reject
Confidence: 2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked. |
This paper introduces Latent Feature Alignment (LFA), a simple method for discovering semantically coherent and interpretable subpopulations directly from face recognition embeddings. The experiments cover celebA and RFW datasets and evaluate the method from multiple perspectives including semantic coherence, interpretability, and bias detection.
1. The automatic discovery of algorithmic bias is socially valuable and critical for building fair AI systems.
2. The method is straightforward, with the iterative approach to discovering semantic directions being easy to understand and implement.
3. The experiments cover celebA and RFW datasets and evaluate the method from multiple perspectives including semantic coherence, interpretability, and bias detection.
1. LFA is essentially a simple iterative nearest neighbor search lacking substantial technical innovation. The entire algorithm reduces to computing average vectors, finding similar samples, and updating directions. The contribution of this paper may not meet the standard of ICLR.
2. The paper only compares against the most basic k-means clustering, lacking comparisons with modern clustering methods and specialized bias detection approaches. Without systematic comparisons to advanced methods, the claimed advantages remain unconvincing.
3. Relying on VLM annotations introduces additional bias and errors. Using potentially biased vision-language models to evaluate system bias severely undermines the credibility of evaluation results. The accuracy of VLMs in handling demographic features also lacks proper validation.
4. Interpretability validation is highly dependent on the quality of a specific generative model (arc2face). This verification approach lacks generality and objectivity, with the biases and limitations of the generative model directly affecting the credibility of validation results.
5. The method lacks validation on other visual recognition tasks. Testing only on face recognition without demonstrating applicability to object recognition, scene classification, or other visual tasks severely limits the method's generalizability.
6. The discovered "biased groups" may simply be clusters of poor-quality data samples rather than genuine algorithmic bias. High FMR could stem from technical factors like image quality and lighting conditions causing embedding noise. The authors fail to effectively distinguish between technical issues and genuine bias problems.
The proposed LFA method lacks substantial technical innovation. The evaluation methodology is flawed, relying on potentially biased VLM annotations and a single generative model (arc2face) for validation, while only comparing against basic k-means clustering without modern alternatives. Additionally, the discovered "biased groups" may represent poor data quality rather than genuine algorithmic bias, as the authors fail to distinguish between technical artifacts and real bias issues. |
Moderately AI-edited |