acceptodds
Under review as a conference paper at ICLR 2027

Population-Level Representation Variability and Triplet-Derived Model-Level Representation Descriptors

Abstract

Independently trained neural networks can develop different representations under the same learning conditions, motivating a population-level view of learned representations. Yet, most analyses focus on pairwise comparisons, limiting characterization of population variability and individual representations. We address both levels with Representational Dispersion (RD), a measure of population-level variability, and the Triplet-Derived Representation Descriptor (TRD), which characterizes each model's representation through relative distances over shared observations. We show that RD captures consistent variability patterns across learning conditions and training stochasticity, while TRD preserves Centered Kernel Alignment (CKA)-based relationships between models despite differences in their representation spaces. We further introduce Compact TRD and Surrogate-Guided Subset Optimization (SGSO), which efficiently optimizes an observation subset under a fixed budget to construct a compact TRD that preserves model-level representation geometry. Together, these tools extend representation analysis from pairwise similarity to population structure and provide a basis for large-scale model characterization, organization, and selection.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Representational Variability, Centered Kernel Alignment, Model Population, Representation Descriptors, Triplet-Based Representation

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_ooSvOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

In this paper, the authors build the theoretical framework for analyzing the impact of different learning conditions to neural networks training through similarity analysis that go beyond the comparison between only two models. To achieve this, the authors introduce Representational Dispersion to measure variability and Triplet-Derived Representation Descriptor (TRD), which characterizes each model’s representation through relative distances over shared observations. Furthermore, the authors propose a Compact version of TRD that can be readily obtained by efficiently optimizing an observation subset under a fixed budget (SGSO).

Critical strengths
  • Extending similarity measurements from two sets of paired representations towards more than two is conceptually interesting
  • Providing an algorithmic approach enabling the selection of a subset of observations that is representative of the set of all observations across the various models is a practically useful idea
Critical weaknesses
  • Despite showcasing interesting methodological ideas, I find the manuscript difficult to read due to phrasing that leads to potentially wrongful interpretations, the use of functions that were not defined before, overwhelming notation and missing intuitive schemes describing the method.
    • The Abstract and Introduction of the paper are somewhat confusing for readers coming from a representation alignment background. It seems as though the paper is proposing new similarity metrics that do not leverage pairwise distances/comparisons, but rather higher order relational structures. For instance:

      Yet, most analyses focus on pairwise comparisons, limiting characterization of population variability and individual representations.

      Centered Kernel Alignment (CKA) (Kornblith et al., 2019), for example, compares two representations but does not directly quantify variability across a population of independently trained models.

      These seem like the authors are proposing a new similarity measure that instead of considering the pairwise distances, as does CKA, it considers another relational structure. This made me very confused while reading the paper for the first time.

    • Algorithm 1 is a bit too informal for scientifical understanding

    • Functions like sim, corr and cos are not defined - making it hard to understand the methological section in 3.3

    • No figure explaining the method makes it more challenging for readers to understand the paper.

    • The pairwise ranking loss in section 3.4 is not defined, making it difficult to analyze the loss

    • I find the experiment section hard to understand, making it difficult for me to accurately review it.

  • A recent published paper utilizing triplet relationships for representation alignment is conceptually very similar to the computation of TRD [1]

[1] Soares, D., Gawade, P., Dittadi, A. & Szczurek, E.. (2026). Scalable and Interpretable Representation Alignment with Ordinal Similarity. Proceedings of the 43rd International Conference on Machine Learning

Questions for authors

How does your TRD differ from the approach in [1]?

How does your work advance the field of representation learning? Which limitations of existing work does your approach address? Why are these limitations important to solve?

Reviewer_ZfNSOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper looks at variability among independently trained neural representations under a fixed learning condition. It proposes Centred Kernel Alignment -based Representational Dispersion (RD) to measure population-level variability, and a Triplet-Derived Representation Descriptor (TRD) that describes each model by relative distances over a shared set of observation triplets. It then introduces Compact TRD (cTRD) with Surrogate-Guided Subset Optimization (SGSO), which searches for a small observation subset whose induced triplet relations preserve the model-level geometry of the full descriptor. Experiments use binary ImageNet-100 domain pairs and multiclass CUB/Cars/ImageNet-100 datasets with ResNet, ViT, and DeiT backbones. The results suggest that RD is stable across training randomness and varies by domain. TRD corresponds well with CKA geometry and the SGSO-selected cTRDs outperforms several baselines.

Critical strengths
  1. The authors ask how variable representations produced by different models are which is an interesting research question.
  2. The study is performed under a systematic control of sources of randomness.
  3. Clear presentation.
Critical weaknesses
  1. The proposed triplet-derived representation descriptor (TRD) is not scalable, and in any case does not mathematically offer anything over the representation dissimilarity matrix (RDM).
  2. The close relationship between representation similarity analysis on Euclidean RDMs and linear CKA has been noted before (Kornblith et al., 2019; Williams et al., 2021; Harvey et al., 2024). In view of weakness 1, the high correspondence between TRD and CKA is expected.
  3. The relationship between RD and CKA is trivial. It appears to be anticorrelated with task difficulty. Is it what it measures, rather than dispersion?
  4. The reported dispersion is likely to be heavily affected by the training regime, which is not justified. The models appear to be undertrained.
  5. CKA estimation is likely to be problematic for the small number of samples used.
  6. There seems to be no downstream use case for the “Large-scale model characterisation, organisation, and selection” capability of the proposed methodology. This raises questions about the practical significance of the work.
Questions for authors
  1. Since TRD is a fixed linear transform of the pairwise-distance matrix, what does TRD offer over the RDM as a descriptor?
  2. Does RD carry information beyond accuracy or R@1?
  3. Are the conclusions unchanged with debiased CKA, given n = 100 observations and up to 2,048 dimensions?
  4. What are the final accuracies of the from-scratch models, and do the RD findings hold for well-trained or pretrained models (e.g. fine-tuned from ImageNet weights)?
  5. What concrete downstream task (model selection, ensembling, property prediction) benefits from TRD?
Reviewer_Hc8MOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper systematically studies full pairwise similarities of models trained with stochasticity. A metric similar to a statistical variance (RD) is introduced. Then, a Triplet-Derived Representation Descriptor (TRD) is introduced, which characterizes each model's representation through relative distances over shared observations. Authors show that for a selected subset of inputs TRD can be similar to CKA, but, contrary to CKA, it is a vector describing a model. These subsets of triplets are transferable to some degree.

Critical strengths
  1. The paper is well written, language is fine.
  2. Experimental section is comprehensive - it is a large-scale evaluation of training randomness.
  3. The idea of TRD is novel.
Critical weaknesses
  1. RD itself can't be considered as a significant contribution - it is just 1 - CKA. Also CKA has lots of limitations - like sensitivity to outliers, CKA is prone to manipulations, etc.
  2. Practical implications of the proposed method are not clear.
  3. In Abstract: "we show that RD captures consistent variability patterns across learning conditions and training stochasticity." But later, experimental evaluation shows that RD ranking might differ for different architectures. There is a contradiction.
  4. The usage of the terms "observer", "observer-pair" is not clear.
Questions for authors
  1. "However, CKA requires direct access to internal representations for the same inputs, limiting its applicability to models or observers whose internal states are inaccessible, such as humans. We therefore examined whether TRD can characterize both pairwise representation relationships and population-level variability."

I don't understand this point. TPD by definition, also needs access to internal representations of a model. So it is not a "black-box" representation comparator.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.