Pith. sign in

REVIEW 3 major objections 4 minor 37 references

Selective Matching Losses -- Not All Scores Are Created Equal

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A matching loss's link slope can be chosen to emphasize chosen score regions; composite Softmax with nonlinear transform $Q(z)$ extends this to multi-class with ranking sensitivity.

desk verdict A genuinely new construction for selective losses with coherent math, but the application-level claims are ahead of the evidence. read the letter →

arxiv 2506.04446 v2 pith:WXW22UAY submitted 2025-06-04 cs.LG

Gil I. Shamir , Manfred K. Warmuth This is my paper
classification cs.LG MSC 68T0568T07
keywords selectivelossesmatchingBregmandivergencelinkfunctionscompositeSoftmaxrankingsensitivityregionlossdesign
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper argues that a wide class of loss functions can be made deliberately selective: by choosing an increasing link function whose slope is steep where accuracy matters and flat where it does not, the loss automatically penalizes errors in focus regions more heavily while discounting nuisance regions. It proves that local sensitivity of such a matching loss is governed by the slope $h'(z)$, and it quantifies, through a bias-underspecification model, how this slope steers an under-specified model to explain high-sensitivity regions at the expense of low-sensitivity ones. For multi-class problems, the paper introduces composite Softmax, in which scores are first passed through a nonlinear log-score transform $Q(z)$; the transform's derivative $q(z)$ acts as a per-score amplification giving both region sensitivity and ranking sensitivity. If correct, this yields a plug-in recipe for emphasizing important score ranges in ranking, distillation, dwell-time prediction, and LLM alignment, while preserving the same gradient structure as classical matching losses.

What carries the argument

The load-bearing object is the link function $h(z)$ (with primitive $H(z)$) in the matching loss $L_m(\hat s,s)=H(\hat s)-H(s)-(\hat s-s)h(s)$. Its slope $h'(z)$ is the local sensitivity, and the bias-underspecification sensitivity (BUST) analysis computes how an underspecified model's optimal prediction shifts toward high-slope regions: the bias sensitivity is proportional to $h'(w_u)/(h(S_M)-h(S_m))$. In the multi-class case the mechanism is composite Softmax: replacing the standard log partition with $H(z)=\log\sum_k e^{Q(z_k)}$ makes each link component $h_k(z)=q(z_k)p_k(z)$, so the scaling $q=Q'$ and the Softmax probability $p_k$ act as two knobs for region and ranking sensitivity. A monotone nondecreasing $q$ (convex $Q$) is the sufficient condition that makes the induced scalar and multi-class losses convex.

What would settle it

Train the same architecture on a ranking or distillation task with standard Softmax and with composite Softmax using a fixed nonlinear $Q(z)$ (say $q(z)=e^z$), keeping all other hyperparameters identical, and measure whether high-score items improve on a metric that weights high relevance (e.g., NDCG) in the way the sensitivity analysis predicts. If the trained model's weights do not shift toward high-sensitivity features as the BUST derivation prescribes, the central claim fails; a cleaner laboratory check is to perturb one observed score by a controlled $\delta$ in a high-sensitivity region and confirm that the loss gradient scales as $h'(\cdot)$ predicts.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is a recipe for designing loss sensitivity rather than inheriting it from a fixed loss. A matching loss is the Bregman divergence $D_H(\hat s,s)$ with gradient $h(\hat s)-h(s)$, where the link $h$ is nondecreasing and $h'(z)$ is the local sensitivity; scalar choices of shifted scaled Sigmoid, exponential, anti-exponential, or $\sinh$ links produce four distinct sensitivity profiles. For multiple classes, the paper constructs composite Softmax probabilities $p_k(z)=e^{Q(z_k)}/\sum_j e^{Q(z_j)}$ and shows the link components become $h_k(z)=q(z_k)p_k(z)$ with $q=Q'$. The resulting multi-class matching loss has $k$-th gradient component $q(\hat s_k)p_k(\hat s)-q(s_k)p_k(s)$, which the authors assert provides flexibility unavailable with standard Softmax: $q$ controls region sensitivity and $p_k$ controls ranking sensitivity, so training can emphasize high scores, low scores, or high norms while remaining a proper Bregman loss.

Load-bearing premise

The claimed link between steeper link slopes and better trained-model behavior rests on the BUST underspecification model, which assumes underspecification acts as a small uniformly distributed interval of feature scores with an added bias; if real network underspecification does not look like that, the practical benefit of selective losses is not established.

Editorial extensions

If this is right

  • Scalar regression and classification losses can be tuned to four sensitivity profiles—low norm, high score, low score, and high norm—by choosing among shifted/scaled Sigmoid, exponential, anti-exponential, and $\sinh$ links.
  • Multi-class losses built from composite Softmax with nonlinear $Q$ provide both region sensitivity (via $q$) and ranking sensitivity (via $p_k$), which standard Softmax cannot supply because of shift invariance.
  • Selective matching losses remain proper scoring rules: at the optimum $h(\hat s)=\mathbb{E}_P h(S)$, so selectivity reweights where model capacity is spent without breaking the expected-link calibration.
  • The BUST result implies that under-specified models resolve ambiguity in favor of high-sensitivity regions, so low-score nuisance examples are discounted without enabling them to corrupt high-score predictions.
  • The paper identifies high-score sensitivity as the useful profile for retrieval, ranking, distillation, and LLM alignment, and capped high-score sensitivity for dwell-time prediction.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial extension: because the construction defines losses by their gradient, it works even when $H$ has no closed form; learned or piecewise links optimized for downstream metrics are a natural avenue the paper does not explore.
  • Editorial extension: the high-norm $\sinh$ link suggests a contrastive loss that suppresses small pairwise score differences and amplifies large ones; the paper mentions ordering robustness but proves no formal noise-robustness guarantee.
  • Editorial extension: swapping standard Softmax for composite Softmax in an existing LLM alignment pipeline is a drop-in test of the framework; the paper states the intended application but provides no empirical comparison, so the practical size of the benefit remains open.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper proposes a framework for designing 'selective' matching losses for scalar and multi-class prediction. A matching loss is defined as a Bregman divergence whose gradient is h(bs)-h(s), where h is a link function; the slope h'(z) is interpreted as local sensitivity. The authors show that shifted/scaled Sigmoid and sinh links produce different region-sensitivity profiles, and they introduce composite Softmax / composite Sigmoid constructions in which the link components factor as q(z_k) p_k(z). Theorems 2-4 give the factorization and sufficient convexity conditions, and extensive appendices catalogue valid and invalid link/scaling choices. The stated motivation is practical: selective losses are claimed to improve dwell-time prediction, retrieval, ranking, distillation, and LLM alignment by emphasizing important score regions and discounting unimportant ones.

Significance. If the construction holds up, the paper offers a principled design recipe for losses with prescribed region and ranking sensitivity, which is a genuinely useful contribution. The Bregman-divergence framing is elegant, and Theorem 2's factorization of the multi-class link into a unary scaling term q(z_k) and a probability term p_k(z) cleanly separates regional and ranking control. The convexity proofs in Theorems 3-4 are self-contained, and the appendices give a helpful catalogue of valid and invalid score transforms. However, the paper's central application claims go well beyond the evidence: there are no experiments, and the bridge from local loss sensitivity to trained-model improvement is an assumption that is not established. The value of the paper at present is as a theoretical/design contribution, not as a demonstrated source of practical gains.

major comments (3)
  1. [Abstract and Section 7] The abstract and conclusions claim 'substantial advantage' over traditional losses for dwell-time prediction, retrieval, ranking, distillation, and LLM alignment, but the manuscript contains no experiments or benchmarks on any of these tasks. Appendix F is a two-dimensional linear-model illustration, and the figures show loss surfaces and gradient curves, not trained-model performance on real or simulated tasks. This is load-bearing because the stated motivation is practical advantage rather than only mathematical expressiveness. I ask the authors to either provide at least one empirical demonstration with appropriate baselines on one or more of the claimed applications, or to rewrite the application claims as hypotheses and scope them explicitly to the proposed loss construction.
  2. [Section 2, Theorem 1 and Proposition 1] The transfer from loss sensitivity to trained-model behavior is not established. Theorem 1's BUST result, Eq. (10), computes the optimal displacement of a single collective score under a uniform distribution over S with one small biased interval; Proposition 1 only states that the minimizer of the expected matching loss is bs = h^{-1}(E_P h(S)). Neither result implies that a gradient-trained neural network with many features, non-uniform score distributions, and non-convex optimization will 'resolve underspecification' in favor of high-sensitivity regions, as asserted in the text around Eq. (10) and Section 7. This is a correctness-risk concern with the paper's central practical claim. Please provide a formal transfer theorem under stated assumptions, an empirical check, or an explicit statement that the improved behavior in real models is an assumption rather than a consequence of the theory.
  3. [Section 5, Eqs. (29)-(30)] The central notions of 'ranking sensitivity' and 'constellation shift invariance' are used informally and are never formally defined. For example, adding a constant to all scores preserves class rankings, but the loss in Eq. (29) is not invariant under this transformation when Q is nonlinear, so the exact intended invariance is unclear. Since the claimed advantage of composite Softmax over per-dimension losses rests on ranking sensitivity, the paper should define these terms precisely (for example, as invariance of the loss or its gradients under specified transformations of the score vector) and state which properties the proposed loss does and does not satisfy.
minor comments (4)
  1. [Section 2, Eq. (2)] The sentence 'The loss is the additional area covered by h(z) from s to bs' should be worded as signed area or as the area between h(z) and the horizontal line h(s) over the interval between s and bs; for bs < s the integrand h(z)-h(s) is non-positive, so the loss is positive but not an area in the ordinary unsigned sense.
  2. [Section 5] The terms 'ranking sensitivity' and 'constellation shift invariance' would benefit from formal definitions; at present the reader must infer them from Figures 4-5 and the surrounding text.
  3. [Appendix G.3] The list of invalid score transforms is useful, but the domain assumptions should be stated explicitly; several entries, such as f(z)=|z|^d, are invalid on certain domains but not on others.
  4. [Appendix E.2.3] The statement that a composite-Softmax CE loss is 'not guaranteed to be convex' would be easier to verify if a concrete non-convex example or a small counterexample plot were included.

Circularity Check

0 steps flagged · score 1.0 of 10

No material circularity: the loss constructions and sensitivity theorems are derived from stated definitions; the practical application claims are unverified but not circular.

full rationale

The paper's derivation chain is self-contained. Matching losses are defined via link functions and Bregman divergences in Equations (1)-(2), and the connection between link slope and local sensitivity follows from the Taylor expansion in Appendix B, not from any fitted quantity. The BUST/BLUST sensitivity model in Definition 1-2 and Theorem 1 is derived from the matching-loss gradient under an explicitly stated underspecification model; it is an analytical consequence of the loss, not an input that is later relabeled as a prediction. Composite Softmax is derived from first principles: Theorem 2 obtains h_k(z) = q(z_k)p_k(z) from the log-partition function H(z) = log(sum e^{Q(z_k)}) purely by the chain rule, and Theorems 3-4 give sufficient convexity conditions using standard composition arguments. No parameter is fitted to a subset of data and then reported as a prediction, and no claimed result is presupposed by its own definition. The self-citations that appear (e.g., SmeLU from Shamir et al. 2020 and Shamir and Lin 2022, shifted Sigmoid from Hristakeva 2010) are used as examples or prior art, not as load-bearing justification for the central construction. The main weakness is evidential rather than circular: the paper claims substantial advantage for dwell-time prediction, retrieval, ranking, distillation, and LLM alignment without experiments on those tasks, and the transfer from loss sensitivity to trained-model behavior rests on the idealized BUST model. Those are correctness or scope concerns, not circularity.

Assumptions & free parameters 0 free parameters · 5 assumptions · 0 invented entities

The framework relies only on standard convex analysis and the definition of matching losses from prior work. No data-fitted constants or new physical entities are introduced. The BUST model is an illustrative modeling assumption, not a fitted component.

assumptions (5)
  • standard math Bregman divergences define matching losses, and the gradient of a matching loss is the difference of the link at predicted and observed scores.
    Used in Section 2, Equation (1)-(2), as the foundation of the loss framework.
  • standard math A monotone non-decreasing link function gives a convex loss.
    Invoked throughout the paper to ensure convexity of the designed losses.
  • standard math The composition of a convex non-decreasing function (Softplus) with a convex Q(z) is convex.
    Used in the proofs of Theorems 3 and 4, citing Boyd and Vandenberghe.
  • domain assumption In the BUST model, the collective feature score w is uniformly distributed over S, and a small interval Su has an added bias d.
    Defined in Section 2 before Theorem 1; this idealized model is used to justify that link slope drives emphasis in training.
  • standard math The optimal prediction under a matching loss satisfies h(bs) = E_P h(S), requiring h to be invertible on the relevant range.
    Proposition 1, used for the proper scoring interpretation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Selective Matching Losses -- Not All Scores Are Created Equal." pith.science (2026). https://pith.science/paper/WXW22UAY

@misc{pith2026250604446,
  author       = {Pith},
  title        = {Pith review of: Selective Matching Losses -- Not All Scores Are Created Equal},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WXW22UAY}},
  note         = {Machine review of arXiv:2506.04446}
}
read the original abstract

Learning systems match predicted scores to observations over some domain. Often, it is critical to produce accurate predictions in some subset (or region) of the domain, yet less important to accurately predict in other regions. We construct selective matching loss functions by design of increasing link functions over score domains. A matching loss is an integral over the link. A link defines loss sensitivity as function of the score, emphasizing high slope high sensitivity regions over flat ones. Loss asymmetry drives a model and resolves its underspecification to predict better in high sensitivity regions where it is more important, and to distinguish between high and low importance regions. A large variety of selective scalar losses can be designed with scaled and shifted Sigmoid and hyperbolic sine links. Their properties, however, do not extend to multi-class. Applying them per dimension lacks ranking sensitivity that assigns importance according to class score ranking. Utilizing composite Softmax functions, we develop a framework for multidimensional selective losses. We overcome limitations of the standard Softmax function, that is good for classification, but not for distinction between adjacent scores. Selective losses have substantial advantage over traditional losses in applications with more important score regions, including dwell-time prediction, retrieval, ranking with either pointwise, contrastive pairwise, or listwise losses, distillation problems, and fine-tuning alignment of Large Language Models (LLMs).

Figures

Figures reproduced from arXiv: 2506.04446 by the authors.

Figure 1
Figure 1. Matching loss as area under the link. A Sigmoid σ(·) link (below) with a SoftPlus primitive can be shifted to vary sensitivity regions. h(z) = σ[α(z − β)] △ = 1 1 + e−α(z−β) , (3) H(z) = 1 α · log  1 + e α(z−β)  . (4) For brevity, x △ = α(z − β) for a β shifted, α scaled function. A Sigmoid induced scalar matching loss is Lm (s, s b ) = 1 α · log  1 + e α(sb−β)  − 1 α · log  1 + e α(s−β)  − (sb− s) · σ [α(s − … view at source ↗
Figure 2
Figure 2. Selective loss sensitivity types. Top: links h(z) and h ′ (z) (shifted/scaled Sigmoid - first three, sinh(z) - last) vs. z. Bottom: Selective losses vs. sb, s ∈ {−3, 0, 3} (−3 red, 0 green, 3 blue). norms. Unless scores are already in a compact domain (e.g., probabilities), a sinh(·) link can be capped in a finite range to avoid instabilities. Piecewise Sensitivity: Piecewise monotonically non-decreasing h(z) with c… view at source ↗
Figure 3
Figure 3. Scaling, composite Sigmoid, link and sensitivity vs. score z, and composite Sigmoid selective losses vs. sb for s ∈ {−3, 0, 3} (−3 red, 0 green, 3 blue) for scaling functions q(z) with low norm, high and low score, and high norm sensitivities. Dashed loss curves show CE losses. 5 Multi-class Selective Matching Losses A multi-class matching loss is defined by a Bregman divergence (DH(·, ·)) Lm(bs, s) △ = DH(bs, s) = … view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Low norm, high score, low score, and high norm sensitivities (left-to-right) multi-class losses vs. sbk sk ∈ {−3, 0, 3} (−3 red, 0 green, 3 blue). Top: hk(z) and diagonal h ′ k(z) (some overlap), middle: scalar losses, bottom: multi-class projection to k, sbj = sj (sol…
Figure 5
Figure 5. Figure 5: Per-dimension loss projections (top), and 2-dimensional loss surfaces (contour; center, surface; bottom) for K = 2 classes vs. sbk, with s1 = 3 and s2 ∈ {1, 3, 5}, for scalar decomposed losses with h(z) = αex (left three columns) and composite Softmax losses with q(z) …

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

37 extracted references · 29 canonical work pages

  1. [1]

    Better click quality on display ads improves the user and advertiser experience, 2015

    Inside AdWords. Better click quality on display ads improves the user and advertiser experience, 2015

    work page 2015

  2. [2]

    Robust bi-tempered logistic loss based on bregman divergences

    Ehsan Amid, Manfred KK Warmuth, Rohan Anil, and Tomer Koren. Robust bi-tempered logistic loss based on bregman divergences. Advances in Neural Information Processing Systems, 32, 2019

    work page 2019

  3. [3]

    Locoprop: Enhancing backprop via local loss optimization

    Ehsan Amid, Rohan Anil, and Manfred Warmuth. Locoprop: Enhancing backprop via local loss optimization. In International Conference on Artificial Intelligence and Statistics, pages 9626--9642. PMLR, 2022

    work page 2022

  4. [4]

    Exponentially many local minima for single neurons

    Peter Auer, Mark Herbster, and Manfred KK Warmuth. Exponentially many local minima for single neurons. Advances in neural information processing systems, 8, 1995

    work page 1995

  5. [5]

    Improving post-click user engagement on native ads via survival analysis

    Nicola Barbieri, Fabrizio Silvestri, and Mounia Lalmas. Improving post-click user engagement on native ads via survival analysis. In Proceedings of the 25th International Conference on World Wide Web, pages 761--770, 2016

    work page 2016

  6. [6]

    Asymmetric loss functions for forecasting in criminal justice settings

    Richard Berk. Asymmetric loss functions for forecasting in criminal justice settings. Journal of Quantitative Criminology, 27: 0 107--123, 2011

    work page 2011

  7. [7]

    Convex Optimization

    Stephen P Boyd and Lieven Vandenberghe. Convex Optimization. Cambridge University Press, 2004

    work page 2004

  8. [8]

    L. M. Bregman. The relaxation method of finding the common point of convex sets and its application to the solution of problems in convex programming. USSR Computational Mathematics and Mathematical Physics, 7 0 (3): 0 200--217, 1967

    work page 1967

Show all 37 references
  1. [9]

    Loss functions for binary class probability estimation and classification: Structure and applications

    Andreas Buja, Werner Stuetzle, and Yi Shen. Loss functions for binary class probability estimation and classification: Structure and applications. Working draft, November, 3: 0 13, 2005

    2005

  2. [10]

    Learning to rank using gradient descent

    Chris Burges, Tal Shaked, Erin Renshaw, Ari Lazier, Matt Deeds, Nicole Hamilton, and Greg Hullender. Learning to rank using gradient descent. In Proceedings of the 22nd international conference on Machine learning, pages 89--96, 2005

    2005

  3. [11]

    From ranknet to lambdarank to lambdamart: An overview

    Christopher JC Burges. From ranknet to lambdarank to lambdamart: An overview. Learning, 11 0 (23-581): 0 81, 2010

    2010

  4. [12]

    Learning to rank: from pairwise approach to listwise approach

    Zhe Cao, Tao Qin, Tie-Yan Liu, Ming-Feng Tsai, and Hang Li. Learning to rank: from pairwise approach to listwise approach. In Proceedings of the 24th international conference on Machine learning, pages 129--136, 2007

    2007

  5. [13]

    An iterative row-action method for interval convex programming

    Yair Censor and Arnold Lent. An iterative row-action method for interval convex programming. Journal of Optimization theory and Applications, 34 0 (3): 0 321--353, 1981

    1981

  6. [14]

    Online learning to rank with top-k feedback

    Sougata Chaudhuri and Ambuj Tewari. Online learning to rank with top-k feedback. Journal of Machine Learning Research, 18 0 (103): 0 1--50, 2017

    2017

  7. [15]

    Christiano, Jan Leike, Tom Brown, Miljan Martic, Shane Legg, and Dario Amodei

    Paul F. Christiano, Jan Leike, Tom Brown, Miljan Martic, Shane Legg, and Dario Amodei. Deep reinforcement learning from human preferences. In Advances in Neural Information Processing Systems, 2017

    2017

  8. [16]

    Support-vector networks

    Corinna Cortes. Support-vector networks. Machine Learning, 1995

    1995

  9. [17]

    Kl divergence for machine learning

    Dibya Ghosh. Kl divergence for machine learning. https://dibyaghosh.com/blog/probability/kldivergence.html, 2023

    2023

  10. [18]

    Worst-case loss bounds for single neurons

    David Helmbold, Jyrki Kivinen, and Manfred KK Warmuth. Worst-case loss bounds for single neurons. Advances in neural information processing systems, 8, 1995

    1995

  11. [19]

    Distilling the knowledge in a neural network

    Geoffrey Hinton. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015

    2015 arXiv

  12. [20]

    Nonlinear Asymmetric Loss with Spectral Decomposition for Glucose Prediction

    Maya H Hristakeva. Nonlinear Asymmetric Loss with Spectral Decomposition for Glucose Prediction. University of California, Santa Cruz, 2010

    2010

  13. [21]

    Robust estimation of a location parameter

    Peter J Huber. Robust estimation of a location parameter. In Breakthroughs in statistics: Methodology and distribution, pages 492--518. Springer, 1992

    1992

  14. [22]

    a rvelin and Jaana Kek \

    Kalervo J \"a rvelin and Jaana Kek \"a l \"a inen. Cumulated gain-based evaluation of ir techniques. ACM Transactions on Information Systems (TOIS), 20 0 (4): 0 422--446, 2002

    2002

  15. [23]

    Relative loss bounds for multidimensional regression problems

    Jyrki Kivinen and Manfred KK Warmuth. Relative loss bounds for multidimensional regression problems. Advances in neural information processing systems, 10, 1997

    1997

  16. [24]

    Learning to rank for information retrieval

    Tie-Yan Liu. Learning to rank for information retrieval. Foundations and Trends in Information Retrieval, 3 0 (3): 0 225--331, 2009

    2009

  17. [25]

    Learning to rank for information retrieval

    Tie-Yan Liu. Learning to rank for information retrieval. Springer, 2011

    2011

  18. [26]

    Manning, Prabhakar Raghavan, and Hinrich Sch \"u tze

    Christopher D. Manning, Prabhakar Raghavan, and Hinrich Sch \"u tze. Introduction to Information Retrieval. Cambridge University Press, 2008

    2008

  19. [27]

    Rectified linear units improve restricted boltzmann machines

    Vinod Nair and Geoffrey E Hinton. Rectified linear units improve restricted boltzmann machines. In ICML, 2010

    2010

  20. [28]

    Supervised learning: No loss no cry

    Richard Nock and Aditya Menon. Supervised learning: No loss no cry. In International Conference on Machine Learning, pages 7370--7380. PMLR, 2020

    2020

  21. [29]

    Direct preference optimization: Your language model is secretly a reward model

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D Manning, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. In Advances in Neural Information Processing Systems, 2023

    2023

  22. [30]

    On ndcg consistency of listwise ranking methods

    Pradeep Ravikumar, Ambuj Tewari, and Eunho Yang. On ndcg consistency of listwise ranking methods. In Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics, pages 618--626. JMLR Workshop and Conference Proceedings, 2011

    2011

  23. [31]

    Composite binary losses

    Mark D Reid and Robert C Williamson. Composite binary losses. The Journal of Machine Learning Research, 11: 0 2387--2422, 2010

    2010

  24. [32]

    Information, divergence and risk for binary experiments

    Mark D Reid and Robert C Williamson. Information, divergence and risk for binary experiments. Journal of Machine Learning Research, 12 0 (3), 2011

    2011

  25. [33]

    Real world large scale recommendation systems reproducibility and smooth activations

    Gil I Shamir and Dong Lin. Real world large scale recommendation systems reproducibility and smooth activations. arXiv preprint arXiv:2202.06499, 2022

    2022 arXiv

  26. [34]

    Smooth activations and reproducibility in deep networks

    Gil I Shamir, Dong Lin, and Lorenzo Coviello. Smooth activations and reproducibility in deep networks. arXiv preprint arXiv:2010.09931, 2020

    2010 arXiv

  27. [35]

    Learning to approximate a bregman divergence

    Ali Siahkamari, Xide Xia, Venkatesh Saligrama, David Casta \ n \'o n, and Brian Kulis. Learning to approximate a bregman divergence. Advances in Neural Information Processing Systems, 33: 0 3603--3612, 2020

    2020

  28. [36]

    Beyond clicks: dwell time for personalization

    Xing Yi, Liangjie Hong, Erheng Zhong, Nanthan Nan Liu, and Suju Rajan. Beyond clicks: dwell time for personalization. In Proceedings of the 8th ACM Conference on Recommender systems, pages 113--120, 2014

    2014

  29. [37]

    Ziegler, Nisan Stiennon, Jeffrey Wu, Tom B

    Daniel M. Ziegler, Nisan Stiennon, Jeffrey Wu, Tom B. Brown, Alec Radford, Dario Amodei, Paul Christiano, and Geoffrey Irving. Fine-tuning language models from human preferences. arXiv preprint arXiv:1909.08593, 2020

    1909 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.