acceptodds
Withdrawn from ICLR 2027

Geometry, Preference, and Shaping: A Unified View of Precision–Recall Objectives

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Region-overlap objectives are usually presented as separate formulas, obscuring which differences change their optimization behavior. We organize a broad class of precision–recall scores through one construction: a weighted quasi-arithmetic mean followed by an outer transformation. Its generator, weight, and outer map specify normalized geometry, preference, and shaping. Dice, IoU, , and Tversky share a harmonic core; arithmetic and geometric aggregation belong to the same power-mean family. The score's derivative ratio determines how it mixes precision and recall gradients, and these coefficients together with the model's metric-gradient Gram matrix determine angular discrepancies. This separates shared geometry from directional equivalence and explains why shaping can preserve each item's direction yet rotate an aggregate update. We characterize compatible ratio fields and recover the generator, preference, and outer shaping of representable scores. Controlled numerical diagnostics validate the identities and constructions. The resulting framework connects objective classification, parameter-gradient interpretation, and inverse design without equating a common representation with identical training dynamics.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Precision–Recall Objectives, Loss Function Design, Gradient Geometry

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_w4YROpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes a unified framework for understanding precision–recall objectives, including Dice, IoU, FβF_\beta, and Tversky, by decomposing them into three components: a generator ϕ\phi that determines the aggregation geometry, a weight ww that controls the relative preference for precision and recall, and an outer transformation ψ\psi that shapes the score. It further analyzes how these objective-level choices interact with model gradients to influence parameter-update directions, particularly showing that transformations that preserve individual gradient directions can still change aggregate directions after reduction. Finally, it characterizes conditions for constructing objectives from desired gradient-ratio fields. The framework is validated through numerical diagnostics, but its practical benefits for real segmentation tasks remain unverified.

Critical strengths

(1) A coherent framework for understanding multiple objectives. The decomposition into geometry, preference, and shaping provides an intuitive way to distinguish why two precision–recall scores differ. It brings Dice, IoU, FβF_\beta, and Tversky into a common mathematical framework.

(2) An elegant result on gradient-direction discrepancies. Proposition 1 derives an exact expression for the angular discrepancy between two objectives using their gradient-ratio difference and the Gram determinant of the component gradients. This clarifies how objective differences and model geometry jointly determine update directions.

Critical weaknesses

(1) Lack of practical guidance for loss design. The paper does not provide clear guidelines on how to leverage the proposed framework to design better training objectives or how to select the components and parameters to improve model performance in practice.

(2) Lack of empirical validation on standard benchmarks. The analysis is primarily supported by theoretical results and numerical diagnostics, without experiments on standard benchmarks to demonstrate whether the proposed insights lead to tangible improvements in real-world tasks.

Questions for authors

See weaknesses.

Reviewer_6LVhOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper focuses on a family of functions SS: a monotonically increasing transformation of a weighted ff-mean of two variables (Eq. 1).

Motivation. The authors’ motivation stems from the binary segmentation of images, which is a per-pixel two-class soft classification of pixels between background and foreground (line 108). Precision PP and recall RR are two standard scores that one aims to maximize. So, when SS is applied to PP and RR, it is assumed that it can be used as a loss function. These are the « region-overlap objectives » as called by the authors.

Results. The paper presents an analytical study of the gradients of SS. No insight is gained from the fact that the variables PP and RR represent precision and recall in classification. Using any other two variables would lead to the same results. The analytical results are compared to empirical results obtained using 8 images of 64 pixels and 12 parameters only, with these images consisting of pure noise.

Critical strengths

A wide variety of performance scores have been defined for classification problems, even when there are only two classes. It seems a good idea to study families of scores that encompass several traditional scores, and in particular those making a tradeoff between precision PP and recall RR.

The use of the PP and RR scores is not limited to the pixel-based binary image segmentation (two-class classification) case mentionned by the authors. These scores also appear in information retrieval problems where true negatives are ignored. They also appear in event detection.

Critical weaknesses

Presentation

  • [W1] Lines 788, 790, 793, and 852 refer to files, but no supplementary material has been submitted. Figure 5, which is mentioned in the main paper, is not in it (the paper must be self-contained).

  • [W2] Some mathematical concepts and symbols are defined well after their first appearance. For example, SRS_R and SPS_P, which appear in Eq. (2), are not defined until l112. The term “shared normalized geometry,” used starting on l50, is not defined until l200. The term “shared directions,” used starting on l51, is not defined until l215. Does “positive partial derivatives” (l43) mean SR>0S_R > 0, SP>0S_P > 0? The term “cocycle” is used 4 times without the cocycle in question being clearly identified. The S~\tilde{S} symbol used l132 is only defined on l246. What is the difference between a dot above a letter (like in Eq. 6) and a prime after it (like in Eq. 8) ?Etc.

  • [W3] In Fig. 1, the terms “orthogonal”, “correlated”, and “near rank one” are undefined.

  • [W4] The context of the work is not clearly specified. From the very first sentence of the abstract, “region-overlap” only makes sense once we learn l107–108 that the focus is on distinguishing between background (BG) and foreground (FG) pixels in images. We can then guess that the overlap is between the predicted FG region and the actual FG region. In computer vision, the problem of distinguishing the BG from the FG is called background subtraction. The authors make no reference to the extensive literature on that subject. See the following papers to get started.

    • T. Bouwmans et al. Deep neural network concepts for background subtraction: A systematic review and comparative evaluation. Neural Networks, 117:8–66, September 2019.
    • Y. Wang et al. CDnet 2014: An expanded change detection benchmark dataset. In IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pp. 393–400, Columbus, Ohio, USA, June 2014.

Informativeness

  • [W5] The links to the literature about the tradeoffs that can be established between precision and recall is missing. This limits the reader's perception of the value added by this paper. For an example of recent paper in that literature, see

    • S. Piérard et al. What is the optimal ranking score between precision and recall? We can always find it and it is rarely F1F_1. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Denver, Colorado, USA, June 2026
  • [W6] The paper does not explain what the best way of designing a score SS for a given application is. In particular, what is the purpose of Section 5.2? Why is it easier to choose a field than to choose a function SS?

  • [W7] The paper does not explain what is the practical impact at the end of the learning of choosing one or another in the SS-family.

Soundness

  • [W8] There is insufficient information to properly define the studied family. Eq. (1): is w∈[0,1]w \in [0,1]? Is ϕ\phi strictly monotonic, as required by an ff-mean (see Kolmogorov, 1930)? Is ψ\psi strictly monotonic so as not to alter the ranking? Are SR>0S_R > 0 and SP>0S_P > 0?

    • A. N. Kolmogorov. Sur la notion de la moyenne. Atti della Academia Nazionale dei Lincei, 12(9):388–391, 1930.
  • [W9] The meaning of PP and RR is not taken into account. For example, the unachievable region in the PR space is ignored. Moreover the tradeoff between PP and RR is also ignored. One cannot assumed that PP and RR are independent variables and that one can improve both of them util reaching the oracle.

    • K. Boyd (2012, December). Unachievable region in precision-recall space and its effect on empirical evaluation. ICML, p. 349.
  • [W10] The authors imply l69 that the ultimate goal is to optimize segmentation accuracy. On the one hand, this is debatable. On the other hand, this score (which does not ignore true negatives) is not part of the family of scores they are studying (all of which completely ignore true negatives). The QQ scores of Eq. (13) appear to be the intersection between the family of scores in Eq (1) and (up to a multiplicative constant) the family of “ranking scores” used in event detection, which includes the accuracy.

    • A. Zakharova et al. Event detection in videos: A framework for the development of new methods. arXiv, abs/2607.04372, 2026.
  • [W11] In binary classification or segmentation, there is no reason to limit oneself to scores of the form shown in Eq 1. Table 1 is, in fact, misleading: only the scores in category A belong to this family. For example, balanced accuracy (l345) assigns a nonzero weight to true negatives, which is impossible to achieve with this family.

  • [W12] The paper does not specify what is expected from a score used as a loss function. Intuitively, we would want the classifier performance to always improve during training. One therefore wants to be able to consider one performance better than another iif the value of the score for the former is greater than the value for the latter. If this is the case for a score SS from family (1), then it is also the case for a score S’S’ obtained by changing ψ\psi (provided that ψ\psi is monotonically increasing). However, changing the ϕ\phi function can break this property. There is no discussion of this in the paper. I recommend consulting recent papers on the subject, such as the following.

    • M. Gösgens et al (2021). Good classification measures and how to find them. Advances in neural information processing systems, 34, 17136-17147.
    • S. Piérard et al (2025, June). Foundations of the theory of performance-based ranking. In CVPR (pp. 14293-14302). IEEE.
Questions for authors
  • [Q1] You use the terms “metric” and “score.” How do you distinguish between these two concepts?

  • [Q2] In equation (2), why do you write SR(SP/SRgP+gR)S_R(S_P/S_R g_P + g_R) insead of SP(SR/SPgR+gP)S_P(S_R/S_P g_R + g_P) or SPgP+SRgRS_P g_P + S_R g_R ? This creates an asymmetry between the roles of RR and PP.

  • [Q3] Where are the quantities defined in equation (6) used?

  • [Q4] Line 188: Why does ϕ\phi need to be strictly increasing? The theory of means only requires that ϕ\phi be strictly monotonic.

  • [Q5] What is MRM_R on line 207?

  • [Q6] What is important? The trajectory in the PR space during training, or the point in the PR space reached at the end of training? All the plots shown in Figure 2 indicate that at the end of training, the upper-right corner (oracle) will be reached. Is this always the case in practice?

  • [Q7] What is the purpose of Proposition 3?

  • [Q8] What is the point of saying “Differentiation proves the result” (l. 272)?

  • [Q9] Why is ψ=tq/q\psi = t^q/q instead of psi=tqpsi = t^q (l. 415)?

  • [Q10] What is the point of stating that ϕ(t)=S(t,t)\phi(t) = S(t, t) in Equation 18? This is always the case. Therefore, you cannot determine phi based on S, since S depends on phi.

  • [Q11] Why did you include the scores for categories B through E in Table 1? They do not belong to the family of scores studied in your work.

  • [Q12] What does a row like the one regarding the Wasserstein Dice score (l. 366) add for the reader? There are a lot of rows like that in the table.

  • [Q13] What is the purpose of Section 6 ?

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

This paper analyzes various precision-recall (P-R) objectives under a unified framework, with a focus on their gradient directions.

Critical strengths

The paper covers a broad range of P-R objectives

Critical weaknesses
  1. The paper mainly compares the gradients of different objectives but does not discuss how these differences affect training or model performance in real tasks.
  2. The paper appears incomplete. In Appendix C, the authors state that "External validation remains unexecuted" and only provide an implementation plan.
  3. Tables 2 and 3 appear to be duplicates.
  4. The paper does not provide clear insights or practical implications from its findings.
Questions for authors

see above

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

Related papers

Discussion (0)

Sign in to comment.