acceptodds
Under review as a conference paper at ICLR 2027

Semantic Coverage Imbalance: A Hidden Representational Bias in Visual Learning

Abstract

Modern visual models achieve strong overall accuracy but fail on specific classes. While prior work attributes these failures to class imbalance or dataset bias, recent work has identified a more fundamental cause, Semantic Coverage Imbalance (SCI), as uneven coverage of class–concept relations in training data. We ask a complementary question: how is uneven semantic coverage reflected in the learned representations and behavior of trained visual models? To address this, we introduce a simple, model-agnostic post-hoc framework that measures semantic representation structure and semantic reliance from concept-based evidence, applicable across datasets, architectures, and concept sources. Across spurious-correlation benchmarks (WaterBirds), attribute-based datasets (CelebA), general object recognition (CIFAR-100), context-shift benchmarks (NICO++), and real-world medical imaging (ISIC-DICM-17K, MILK10k), we find that the model-level effects of uneven semantic coverage vary across architectures and domains, are associated with worst-class performance and generalization gaps, and are not explained by class imbalance or dataset-level bias alone. Under shortcut and context shift, failures depend on which semantic concepts a model relies on, not simply on the overall degree of semantic reliance. We further show that post-hoc analysis can identify unstable semantic dependencies that provide useful intervention targets. In medical imaging, however, clinically meaningful concepts remain comparatively stable across domains, suggesting that robust prediction depends on relying on the right semantics rather than uniformly reducing semantic reliance. Overall, our framework offers a principled approach for analyzing and diagnosing SCI as a representation-level bias in visual learning settings.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Representation analysis, semantic representations, spurious correlations, model diagnostics

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_3VEMOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper presents a post-hoc diagnostic of how Semantic Coverage Imbalance (SCI), introduced in SemCovNet (https://arxiv.org/abs/2602.16917), is reflected in trained image classifiers. It aims to answer whether per-class concept distribution can explain per-class failure beyond class imbalance A concept function maps images to concept scores z(x), using a linear probe on annotated attributes, CLIP image–text similarity (CIFAR-100), or MONET descriptors (medical data). From class-conditional means of z, the paper defines Semantic Imbalance (variance of class means), Semantic Reliance (concept-score gap between correct and incorrect predictions) and a Unified Reliance Score (per-concept reliance weighted by the source–target change in mean concept score). These are correlated with worst-class performance and shift degradation across 4 backbones on spurious-correlation and context-shift datasets (WaterBirds, CelebA, NICO++), CIFAR-100 and two dermoscopy datasets, and used to guide a feature-attenuation intervention compared with Deep Feature Reweighting (DFR; last-layer retraining on group-balanced data).

Critical strengths
  • The problem is relevant.
  • The definitions of main metrics are easy to follow.
  • experiments are conducted on various domains
Critical weaknesses

Novelty:

  • The paper seems rather incremental over SemCovNet, which is under-credited throughout. The intro frames SemCovNet as merely "training-time" - the novelty compared to it needs to be stated more precisely.
  • The concept evidence metric is the same and some of the post-hoc analysis overlaps. The newly introduced measures are simple first-order moments of the concept scores, given new names.
  • SR (concept/attribute evidence between correctly and incorrectly classified samples) is a widely used concept from related literature on bias, given a new name. The works are not cited. For example: B2T (Kim et al., "Discovering and Mitigating Visual Biases through Keyword Explanation", CVPR 2024), Spuriosity Rankings (Moayeri et al., NeurIPS 2023), CB-SLICE (Konforti et al., ICML 2026). Overall these should be cited and discussed
  • the related work discussion overall omits a lot of important literature, as evidenced above. Also, I believe it is an essential part of the paper and should not be placed in the supplementary.
  • the propositions add little beyond their assumptions. No experiment depends on ie 2.1 and 2.3

Clarity:

  • . L51–52 adopts SCI from SemCovNet as uneven class–concept coverage in the training data. Fig. 4 then labels SI, SR and SI×SR as "SCI variables" / "SCI (SI+SR+SI×SR)". The paper never measures training-data coverage, nor relates it to SI/SR, so the claim to diagnose SCI rests on the name. The Reproducibility Statement promises "coverage computation" and "controlled coverage interventions", neither of which appears in the paper.

Methodology and correctness:

I have many methodological concerns with the work. Some examples are listed here, I could go on.

  • SR is correlational only, there is no evidence the concepts are actually used by the classifier. See for example "Representation Is Not Reliance: Concept-Based Causal Diagnostics of Shortcut Learning in Vision" (ICML 2026 Mech Interp Workshop).
  • App. F.1 claims that prior concept methods do not characterize "how concept usage varies across classes within a model"; Spuriosity Rankings and B2T do that.
  • the framing rests on “modern visual models” and proceeds to cite mostly 2019-2020 work. Moderen encoders (DINo family, SigLIP2, …) are missing.
  • SI is not model-specific in half the settings: ISI is identical across the 4 architectures. This means that z(x) there is the dataset-provided MONET score, not g(f_θ(x)). This contradicts the "representation-level" claim. CIFAR-100 computes z from the CLIP embeddings, which is also independent of the evaluated models.
  • some classifiers predict worst-class F1 = 0; which is essentially a degenerate classifier predicting the majority class. For such a model, it is unclear how SR is even computed as µ+/µ- are undefined. SImilarly frozen CLIP + linear head is broken on WaterBirds. Accuracy 78.4, worst-class F1 0.062, worst-acc 3.2%. This signals an udnertrained or degenerate baseline again.
  • SI is pooled across datasets, but it scales with the number of concepts, so it is nto comparable.
  • Wtaerbirds experiments ios circular: K=2 concepts = land/water, which is the spurious attribute. SR then measures whether background evidence differs between correct and incorrect samples, which is the definition of the shortcut failure. More generally, concept selection determines SI, SR and URS, yet the concept lists per dataset are given only in App. G (Tab. 7), and the main text never discusses how results depend on the vocabulary.

Presentation

  • There are broken references in intro, a “Definition ??” in App. D, undefined or overloaded notation.
  • Fig. 1 draws the prediction arrow from the semantic evidence z into h_θl. In the method, h_θ takes f_θ(x) and z is a side read-out. Draw z as a branch off f_θ(x).
  • The input requirements are only clear from App. G
  • App. B is called "Extended related work" but there is no non-extended version.
  • Intro, para 2 repeats the dataset-bias and spurious-correlation citations already given in para 1
  • Calling CIFAR-100 (32×32) "general object recognition" is an overstatement to me
Questions for authors

The main concerns go beyond what is resolvable with the rebuttal scope and would need a substantial paper revision. The following nevertheless have the highest chance of changing my assessment: Clarify the “diagnostic is a model property” claim Clarify the difference over prior work mentioned above.

Reviewer_Yr5bOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper investigates how semantic imbalance in training data affects learned visual representations. The authors propose a post-hoc analysis framework based on concept evidence. Specifically, the authors introduce Semantic Imbalance (SI) to quantify variations in class-level semantic profiles, Semantic Reliance (SR) to measure differences in concept evidence between correct and incorrect predictions, and Unified Reliance Score (URS) to assess reliance on domain-varying concepts. Experiments span multiple vision architectures and datasets, including medical imaging and distribution-shift benchmarks.

Critical strengths
  • The problem of concept learning (class-level semantic evidence) is interesting. Understanding how class-level semantic evidence relates to model failures is important for visual learning and XAI. The experiments span multiple architectures, datasets, and concept sources, providing useful empirical insights.
Critical weaknesses
  • The proposed metrics are not fully convincing. SI in Eq (4) measures variation in class-level concept means, which may simply reflect class separation but not coverage imbalance. SR measures differences in concept evidence between correct and incorrect predictions, but does not establish actual model reliance on these concepts. Can the authors clarify and provide more intuition or evidence to show that these metrics capture training-data semantic coverage imbalance instead of general representation properties or prediction accuracy?

  • Proposition 2.2 bounds shift-induced loss changes using URS, but the proof in appendix C.2 assumes that the expected absolute representation perturbation is bounded by the difference between source and target concept means, which does not generally hold. Moreover, the derivation converts a first-order approximation into an inequality without bounding the approximation error or properly handling the absolute sensitivity terms. These gaps weaken the theoretical justification for linking URS to distribution-shift vulnerability.

  • The experiments do not directly validate the connection between training-data SCI and the proposed metrics. Most results are based on correlations between SI/SR and model performance across existing datasets and architectures. I suggest the authors to provide a controlled experiment that varies class–concept coverage while keeping class frequencies fixed to establish whether SI and SR actually capture the effects of SCI.

Questions for authors
  • As mentioned in Weakness 1, can the authors directly relate SI and SR to measured training-data SCI?
Reviewer_PaoGOpenReview ↗
Rating
4
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper, focusing on vision models and datasets, proposes metrics (SI, SR, URS) in a post-hoc model agnostic framework to evaluate uneven semantic coverage. They perform experiments utilizing said metrics to identify findings downstream, such as SR correlation with performance and Shift vulnerability. Some important findings are:

  • Robustness depends on the nature of reliance of the model on semantic concepts and their respective nature
  • semantic representation structure can be clearly captured and differentiated between datasets and models
  • said semantic represenation structure can characterize and be used to explain model level differences not captured by more conventional measurements such as class imbalance
Critical strengths
  • Very well written, mostly easy to read and understand
  • Simple yet useful and widely applicable idea
  • Detailed, fleshed out appendix
Critical weaknesses

W1. The paper often uses varying terminology sometimes interchangeably which may lead to confusion. For example, in 3.4, shift drop, shift degradation, performance drop, performance degradation, all have very similar meaning and usage. This may be easily inferred but because this paper has a tendency of using many terms including ones it has introduced eg. Semantic Reliability, Semantic Imbalance, I had to stop and check to confirm. This is one specific example

W2. Some ablations seem like they may be important, as further highlighted in the Questions section

Questions for authors

Q1. It seems as if each of the datasets have their own specific concept sources and this disparity is not addressed or ablated. As this work is highly dependant on concept labels, a very obvious question is how much the concept source affects these results. For example if the CLIP based concept source is used as an alternate concept source for one of the datasets with human-annotated attributes, how significantly will the results change, if at all?

Q2: Similarly it appears that in section 3.4, that each dataset is evaluated in a different specific setting. Is there a reason that ablations were not performed applying these settings across datasets so that the role of dataset vs role can be more clearly disentangled?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.