Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

Regularity-Aware Stochastic MGDA with Adaptive Conflict-Avoidant Update Direction Control

T0 review · 3 major / 4 minor · reviewed 2026-08-01 · deepseek-v4-flash

Pith's one-line read Exploiting a hidden Lipschitz regularity speeds stochastic multi-objective learning from O~(T^-1/4) to O~(T^-1/2).

desk verdict The continuity results are real, but the advertised rate improvement comes from the fixed-scalarization fallback, not from the Lipschitz analysis. read the letter →

arxiv 2607.15412 v1 pith:L74XQ4NW submitted 2026-07-16 cs.LG math.OC

classification cs.LGmath.OC MSC 90C2990C1565K05
keywords multi-objectivelearningmulti-gradientdescentconflictavoidanceHöldercontinuityLipschitzstochasticoptimizationParetostationaritymulti-task
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

This paper pinpoints why vanilla stochastic multi-gradient descent (MGDA) converges slowly: the conflict-avoidant update direction is only 1/2-Hölder continuous in the gradient matrix, so mini-batch noise gets amplified. The paper shows that under a nondegeneracy condition—the simplex subproblem being strongly convex along its tangent space—the direction becomes Lipschitz continuous. It then builds MoRe, an algorithm that uses the MGDA direction in regular regimes and falls back to a fixed scalarization weight in degenerate ones. The result is an improvement in the nonconvex Pareto-stationarity rate from O~(T^{-1/4}) to O~(T^{-1/2}), together with per-iterate guarantees that the update direction stays close to the full-batch conflict-avoidant direction.

What carries the argument

The central object is the conflict-avoidant direction mapping d_Q = -Q λ*_Q, where λ*_Q solves min_{λ∈∆_M} ||Qλ||. The key mechanism is the µ-nondegeneracy condition (Definition 2): with U spanning the simplex tangent space 1^⊥, the matrix U^T Q^T Q U must have smallest eigenvalue ≥ µ, giving strong convexity of the subproblem on the affine hull of the simplex. This forces a unique MGDA weight and yields the Lipschitz bound ||d_{Q1} - d_{Q2}|| ≤ (1 + 2B²/µ) ||Q1 - Q2|| (Lemma 1); absent this condition, only the sharp 1/2-Hölder bound holds (Lemma 2). MoRe (Algorithm 1) turns this into a rule: when µ_min(U^T Q_t^T Q_t U) < µ_t, it replaces the MGDA weight with a fixed scalarization λ_0, prote

What would settle it

A concrete check: implement MoRe on a two-objective quadratic problem whose Hessians are chosen so that the nondegeneracy parameter at the iterates stays above a positive constant, and measure E[min_{t≤T} R_S(x_t)] under the prescribed schedules (|Z_t|=Θ(t+1), α=Θ(T^{-1/2}), µ_t=Θ(1)). If the empirical rate visibly deviates from O(T^{-1/2}) for large T, the claim is false. Conversely, on a problem engineered to approach degeneracy (where the gradients of the two objectives become nearly parallel at the Pareto front), verify that MoRe's rate degrades toward O(T^{-1/4}) as the theory predicts.

Watch

Extended reading notes

Core claim

The paper's central discovery is a dichotomy in the continuity of the conflict-avoidant direction map Q ↦ d_Q, where d_Q = -Q λ*_Q is the minimum-norm convex combination of the objective gradients. On any bounded set of gradient matrices, the map is 1/2-Hölder continuous and the exponent 1/2 is sharp (Lemma 2); no higher-order Hölder bound holds in the worst case. But if the reduced curvature matrix U^T Q^T Q U is uniformly lower bounded by µ > 0 (Definition 2), the map is Lipschitz with constant 1 + 2B²/µ (Lemma 1). This regularity gap explains why stochastic MGDA suffers a suboptimal rate, and motivates MoRe, which tests the curvature condition at each step and switches to fixed-scalarizat

Load-bearing premise

The load-bearing premise is Assumption 2: each mini-batch gradient must be unbiased with variance bounded by σ²/|Z_t|, together with Assumption 4's uniform per-sample gradient bound; if gradient noise is heavy-tailed or gradients are unbounded, the Lipschitz-based error control and the improved O~(T^{-1/2}) rate collapse.

Editorial extensions

If this is right

  • If the theorem is correct, stochastic multi-objective optimization can match the iteration complexity of single-objective nonconvex SGD (O~(T^{-1/2})), while keeping the conflict-avoidant property of MGDA.
  • The per-iterate CA direction distance guarantee gives a quantitative promise that MoRe's updates track the full-batch conflict-avoidant direction, decaying as O(1/log T) under the recommended schedule.
  • The threshold µ_t creates an explicit trade-off: larger thresholds give faster convergence, smaller thresholds give better conflict-avoidance; setting µ_t ≈ t^{-1/3} optimizes the CA distance at O(t^{-1/3}).
  • The sharpness result (Lemma 2) implies that without a regularity condition or a fallback strategy, no stochastic MGDA variant can achieve linear noise propagation, so the suboptimality of vanilla SMG is intrinsic, not an artifact of the analysis.

Reading between the lines

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

  • Editorial inference: the same regularity-switch principle could improve other nonlinear gradient-aggregation methods (e.g., gradient projection or direction-averaging schemes) that suffer from Hölder-type sensitivity to input noise; the curvature diagnostic is general.
  • Editorial inference: the requirement of linearly growing batch sizes is a practical cost; combining MoRe with variance reduction (e.g., multi-objective SVRG) may preserve the O(T^{-1/2}) rate with constant mini-batches, but the paper does not analyze this.
  • Editorial inference: the theory concerns the empirical objective; an extension to population-level guarantees (excess risk) would be needed to certify the observed multi-task learning gains beyond the training set.
  • Editorial inference: in degenerate regions, MoRe effectively abandons conflict avoidance to preserve stability; an alternative—detecting degeneracy and then increasing batch size—might keep CA behavior without sacrificing the rate, which is a testable variant.
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 studies the continuity of the MGDA conflict-avoidant (CA) direction map Q -> d_Q in stochastic multi-objective learning. It proves that the map is 1/2-Hölder continuous on bounded sets, with a sharpness example, and that under a μ-nondegeneracy condition (Definition 2) the map is Lipschitz continuous (Lemma 1). Based on this, the authors propose the MoRe algorithm, which switches between the stochastic MGDA direction and a fixed scalarization weight depending on a measured nondegeneracy threshold. Theorem 1 claims a nonconvex Pareto-stationarity rate of O(T^{-1/2}) (tilde-O with a logarithmic threshold schedule) under growing mini-batches |Z_t| = Θ(t+1) and step size α = Θ(T^{-1/2}), compared with the known tilde-O(T^{-1/4}) rate for SMG. Theorem 2 gives a per-iterate CA-direction-distance bound for M=2. Experiments on Office-Home are reported in Table II and Figure 3.

Significance. The continuity analysis in Lemmas 1 and 2 is clean and the sharpness counterexample in Appendix B is explicit and convincing. If the claimed rate were attributable to the Lipschitz mechanism, the paper would be a substantial advance over stochastic MGDA. However, the central attribution is not supported: Theorem 1's convergence rate can be obtained already by the fixed-scalarization fallback in Eq. (70), without any use of Lemma 1, the threshold μ_t, or the switching rule. Thus the claimed rate improvement over SMG does not demonstrate the value of the regularity-aware CA mechanism. The per-iterate CA-distance guarantee in Theorem 2 is a genuine separate contribution, but it is restricted to M=2. The paper's contribution is therefore best described as a fine-grained continuity analysis plus a provably stable fallback schedule, rather than a rate improvement caused by exploiting Lipschitz continuity.

major comments (3)
  1. [Theorem 1 / Appendix C, Eq. (70)] The claimed convergence-rate improvement is not attributable to the Lipschitz continuity of the CA direction. In Case 2 of the proof, Eq. (70) is exactly the standard scalarized SGD inequality for the fixed weight λ0: α_t R_S(x_t) ≤ λ0^T F_S(x_t) − λ0^T F_S(x_{t+1}) + α_t ℓ_f ||Q_t − ∇F_S(x_t)||_F + O(α_t^2). Summing Eq. (70) over t under |Z_t|=Θ(t+1) and α=Θ(T^{-1/2}) yields E[min_t R_S(x_t)] = O(T^{-1/2}) without ever invoking Lemma 1, the threshold μ_t, or the switching rule. Since R_S(x_t) ≤ ||∇F_S(x_t)λ0||^2, an algorithm that always sets λ_t = λ0 attains the same rate. Therefore the O(T^{-1/2}) rate in Theorem 1 does not support the abstract's claim that exploiting Lipschitz continuity is what enables the improvement over SMG. The paper should either reframe the contribution, or provide a setting where fixed scalarization cannot achieve the rate and the switching mechanism is essen
  2. [Section IV-B / Figure 3] The empirical section does not discriminate between the proposed mechanism and the trivial always-fallback baseline. Figure 3 only reports the stationarity measure R_S(x_t) with a 1/sqrt(T) reference line; as shown in the major comment above, the always-fixed-scalarization algorithm is also predicted to decay like 1/sqrt(T) under the same schedules. To support the regularity-aware contribution, the experiments should compare MoRe against (a) SMG under the same batch and step schedules, and (b) the always-fallback baseline λ_t = λ0. Reporting the fraction of iterations where the regular branch is used and the per-iterate CA distance would also help verify the switching mechanism.
  3. [Abstract and Conclusion] The abstract and conclusion state that the method 'exploits the Lipschitz continuity of the CA direction when the subproblem is regular' and thereby 'improves the convergence rate of SMG from tilde-O(T^{-1/4}) to tilde-O(T^{-1/2})'. Given Eq. (70), this causal statement is not established. The rate improvement over SMG is a property of the growing-batch-size fallback to a fixed scalarization, not of the Lipschitz regularity of the CA map. The narrative should be revised so that the convergence-rate claim is decoupled from the Lipschitz-continuity insight; otherwise the central claimed contribution is overstated.
minor comments (4)
  1. [Assumption 2] The variance bound E||q_{Z_t,m}(x_t) − ∇f_{S,m}(x_t)||^2 ≤ σ^2/|Z_t| is load-bearing for Theorem 1. The Office-Home experiments do not verify this distributional assumption. A brief discussion of the limitation would be appropriate.
  2. [Theorem 2] The per-iterate CA-distance result is restricted to M=2, and the conclusion acknowledges this. This is fine, but since the paper's broader tables (Table I) highlight only the general convergence rate, the M=2 restriction should be stated more prominently in the abstract or contribution list.
  3. [Table II] The Office-Home results are reported without standard deviations or multiple-seed information, so it is hard to judge whether the reported improvement over baselines is significant.
  4. [Appendix C, Eq. (63)] In the regular-branch bound, the indicator 1_{G_t} is dropped to give Eq. (63). This is valid because the term is nonnegative, but the resulting bound is loose. Consider stating explicitly that the price is paid later via the bad-event term.

Circularity Check

0 steps flagged · score 1.0 of 10

Derivation is self-contained; no circularity. The rate theorem is proved from stated assumptions, and self-citations to [2] are not load-bearing.

full rationale

The central derivation chain is not circular. Lemma 1 proves Lipschitz continuity of the CA-direction map from the µ-nondegeneracy condition (Definition 2) using strong convexity of the reduced simplex subproblem; it does not assume the target Lipschitz bound. Lemma 2 proves the 1/2-Hölder bound in-line and establishes sharpness with an explicit two-matrix counterexample whose reduced curvature degenerates as sin^2(t), so it does not rely on the present paper's conclusions. Theorem 1 bounds the Pareto-stationarity measure through a descent inequality on the fixed scalarization λ0, controlling the stochastic direction error with Assumption 2's variance bound and the Lipschitz/bad-event decomposition; the final rate O(T^{-1/2}) under |Z_t|=Θ(t+1), α=Θ(T^{-1/2}) is obtained by summing deterministic inequality terms, not by fitting a constant to data. The proof is valid even if the fixed-scalarization fallback alone would yield a similar rate, as seen in Eq. (70); that is an attribution/ablation concern about which mechanism drives the rate, not a circularity, because the theorem does not assume the CA-direction Lipschitz bound as an input. Theorem 2's per-iterate CA-distance bound follows from a branch decomposition and the M=2 identity µ_min(U^T Q^T Q U)=1/2||q_{Zt,1}-q_{Zt,2}||^2; the manuscript explicitly states the M=2 limitation. Self-citations to [2] (co-authored by L. Chen) are used for definitions of CA-direction distance, baseline method comparisons, and prior rates; the previously established 1/2-Hölder bound is re-derived in Lemma 2 rather than imported, and the claimed O(T^{-1/4}) SMG baseline is an external comparison, not an input to the proof. No parameter is fitted to the target stationarity measure, and no result reduces to its own inputs by construction. Overall, no significant circularity; score 1 reflects only the minor self-citation presence, which is not load-bearing.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

The central convergence/CA-distance theorems rest on four standard stochastic-optimization assumptions plus the paper-specific μ-nondegeneracy condition and exact subproblem solves. The threshold schedule and fallback weight are user-chosen hyperparameters that control the convergence/conflict-avoidance trade-off. No new physical or mathematical entities are postulated.

free parameters (3)
  • Regularity threshold schedule μ_t = Θ(1) or Θ(1/log(e+t)); numerical constants unspecified
    The convergence rate and CA-distance trade-off in Theorems 1–2 depend on this schedule; it is chosen by the user, not fitted to data.
  • Fallback scalarization weight λ_0 = Uniform 1/M in experiments; otherwise any fixed simplex vector
    Determines the fixed scalarization used on the degenerate branch and hence which Pareto-stationary point the fallback updates target.
  • Step size α and batch-size schedule |Z_t| = α = Θ(T^{-1/2}), |Z_t| = Θ(t+1)
    These schedules balance the terms in Eq. (19) to obtain the stated T^{-1/2} rate; load-bearing for the theorem, though standard in stochastic optimization.
assumptions (6)
  • domain assumption Assumption 1: each empirical objective f_{S,m} is bounded below
    Used to bound the telescoping term E[λ_0^T F_S(x_0) − λ_0^T F_S(x_T)] in Theorem 1 (Eq. 82).
  • domain assumption Assumption 2: unbiased mini-batch gradients with variance ≤ σ²/|Z_t|
    Load-bearing for all stochastic error controls, including the bad-event bound Eq. (54) and the Hölder-term bound Eq. (78).
  • domain assumption Assumption 3: every fixed scalarization λ^T F_S is ℓ_{f,1}-smooth
    Provides the descent inequality used in both branches of the proof of Theorem 1 (Eqs. 55, 66).
  • domain assumption Assumption 4: bounded per-sample gradients ∥∇f_{z,m}(x)∥ ≤ ℓ_f
    Used to bound ∥Q_t∥_F, ∥∇F_S(x_t)∥_F, and the Lipschitz/Hölder constants throughout the analysis.
  • ad hoc to paper μ-nondegeneracy of the stochastic gradient matrix (Definition 2)
    New structural condition introduced by this paper; it is the key regime in which Lemma 1 gives Lipschitz continuity and in which Algorithm 1 uses the MGDA direction.
  • domain assumption Exact solution of the simplex subproblem (14) at every regular-branch iteration
    Algorithm 1 line 6 asserts λ_t = argmin ∥Q_t λ∥²; the error analysis assumes d_t = −Q_t λ_t is exactly the min-norm direction. Inexact QP solves are not analyzed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Regularity-Aware Stochastic MGDA with Adaptive Conflict-Avoidant Update Direction Control." pith.science (2026). https://pith.science/paper/L74XQ4NW

@misc{pith2026260715412,
  author       = {Pith},
  title        = {Pith review of: Regularity-Aware Stochastic MGDA with Adaptive Conflict-Avoidant Update Direction Control},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/L74XQ4NW}},
  note         = {Machine review of arXiv:2607.15412}
}
abstract

Multi-objective learning (MOL) aims to optimize multiple objectives simultaneously. The multi-gradient descent algorithm (MGDA) is a workhorse that iteratively updates along a common descent or conflict-avoidant (CA) direction across objectives. In stochastic settings, however, the vanilla stochastic MGDA method, SMG, lacks a fast convergence rate because mini-batch sampling introduces noise in the gradients. This causes bias in the update direction, which is controlled by the CA direction continuity. In this paper, we show that the CA direction is $1/2$-Holder continuous with respect to the Jacobian matrix, and the exponent $1/2$ cannot be improved in the worst case. This leads to a suboptimal convergence rate for vanilla stochastic MGDA in prior works. Nevertheless, under additional regularity conditions, we show this can be improved to Lipschitz continuity. Based on this insight, we propose a stochastic multi-objective regularity-aware (MoRe) method that exploits the Lipschitz continuity of the CA direction when the subproblem is regular, and switches to a fixed scalarization weight otherwise. Intuitively, the proposed algorithm employs CA direction update when the gradient conflict is large, and linear scalarization update otherwise. Theoretically, our method improves the convergence rate of SMG in the nonconvex setting from $\widetilde{\mathcal O}(T^{-1/4})$ to $\widetilde{\mathcal O}(T^{-1/2})$, where $\widetilde{\mathcal O}(\cdot)$ hides logarithmic factors. Meanwhile, we also establish the per-iterate conflict-avoidance guarantees. Empirically, experiments demonstrate its effectiveness in multi-task performance and verify convergence behavior consistent with the established theoretical rate.

Figures

Figures reproduced from arXiv: 2607.15412 by the authors.

Figure 1
Figure 1. Flowchart of the proposed MoRe algorithm. The illustrations show [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Rate trade-off under the choice µt = Θ((t + 1)−γ), 0 < γ < 1/2, with linearly growing mini-batch size |Zt| = Θ(t + 1) and step size α = Θ(T −1/2 ). MoRe achieves CA distance exponent min{γ, 1 − 2γ} and convergence exponent 1/2 − γ. The additional term 1/(|Zt|µ 3/2 t ) in Theorem 1 contributes exponent 1 − 3γ/2, which is larger than 1/2 − γ for the plotted range and therefore is not rate-limiting. The SMG rates are s… view at source ↗
Figure 3
Figure 3. Convergence curve of the MoRe algorithm on Office-Home. The [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Improved Convergence Rate for Stochastic Multi-Gradient Descent: A Proof Discovered with AI

    math.OC 2026-07 accept novelty 6.0 of 10

    Vanilla stochastic multi-gradient descent achieves Õ(T^{-1}) squared Pareto-stationarity under linearly growing mini-batches, improving the prior Õ(T^{-1/4}) bound.

Reference graph

Works this paper leans on

23 extracted references · cited by 1 Pith paper

  1. [2]

    Three-way trade-off in multi-objective learning: Optimization, generalization and conflict- avoidance,

    L. Chen, H. Fernando, Y . Ying, and T. Chen, “Three-way trade-off in multi-objective learning: Optimization, generalization and conflict- avoidance,”Journal of Machine Learning Research, vol. 25, no. 193, pp. 1–53, 2024

  2. [12]

    The multiobjective steepest descent direction is not Lipschitz continuous, but is H ¨older continuous,

    B. F. Svaiter, “The multiobjective steepest descent direction is not Lipschitz continuous, but is H ¨older continuous,”Operations Research Letters, vol. 46, no. 4, pp. 430–433, 2018

  3. [1]

    The stochastic multi-gradient algorithm for multi-objective optimization and its application to supervised machine learning,

    S. Liu and L. N. Vicente, “The stochastic multi-gradient algorithm for multi-objective optimization and its application to supervised machine learning,”Annals of Operations Research, vol. 339, no. 3, pp. 1119– 1148, 2024

  4. [3]

    Multi-task learning as multi-objective op- timization,

    O. Sener and V . Koltun, “Multi-task learning as multi-objective op- timization,” inAdvances in Neural Information Processing Systems, vol. 31, 2018

  5. [4]

    Multi-objective meta learning,

    F. Ye, B. Lin, Z. Yue, P. Guo, Q. Xiao, and Y . Zhang, “Multi-objective meta learning,” inAdvances in Neural Information Processing Systems, vol. 34, 2021, pp. 21 338–21 351

  6. [5]

    Learning with limited samples: Meta-learning and applications to com- munication systems,

    L. Chen, S. T. Jose, I. Nikoloska, S. Park, T. Chen, and O. Simeone, “Learning with limited samples: Meta-learning and applications to com- munication systems,”Foundations and Trends® in Signal Processing, vol. 17, no. 2, pp. 79–208, 2023

  7. [6]

    Fairness constraints: Mechanisms for fair classification,

    M. B. Zafar, I. Valera, M. G. Rogriguez, and K. P. Gummadi, “Fairness constraints: Mechanisms for fair classification,” inProceedings of the 20th International Conference on Artificial Intelligence and Statistics, ser. Proceedings of Machine Learning Research, vol. 54, 20–22 Apr 2017, pp. 962–970

  8. [7]

    Algorithms for multicriterion optimization,

    H. Mukai, “Algorithms for multicriterion optimization,”IEEE Transac- tions on Automatic Control, vol. 25, no. 2, pp. 177–186, 1980

Show all 23 references
  1. [8]

    Conflict-Averse Gradient Descent for Multi-task Learning,

    B. Liu, X. Liu, X. Jin, P. Stone, and Q. Liu, “Conflict-Averse Gradient Descent for Multi-task Learning,” inProc. Advances Neural Inf. Process. Syst., vol. 34, virtual, Dec. 2021, pp. 18 878–18 890. PREPRINT, JULY 2026 7

  2. [9]

    Steepest descent methods for multicriteria optimization,

    J. Fliege and B. F. Svaiter, “Steepest descent methods for multicriteria optimization,”Mathematical methods of operations research, vol. 51, pp. 479–494, 2000

  3. [10]

    Multiple-gradient descent algorithm (mgda) for multi- objective optimization,

    J.-A. D ´esid´eri, “Multiple-gradient descent algorithm (mgda) for multi- objective optimization,”Comptes Rendus. Math ´ematique, vol. 350, no. 5-6, pp. 313–318, 2012

  4. [11]

    Mitigating gradient bias in multi-objective learning: A prov- ably convergent approach,

    H. Fernando, H. Shen, M. Liu, S. Chaudhury, K. Murugesan, and T. Chen, “Mitigating gradient bias in multi-objective learning: A prov- ably convergent approach,” inProc. Int. Conf. Learn. Representations, Kigali, Rwanda, May 2023

  5. [13]

    Gradient Surgery for Multi-Task Learning,

    T. Yu, S. Kumar, A. Gupta, S. Levine, K. Hausman, and C. Finn, “Gradient Surgery for Multi-Task Learning,” inProc. Advances Neural Inf. Process. Syst., vol. 33, virtual, Dec. 2020, pp. 5824–5836

  6. [14]

    On the convergence of stochastic multi-objective gradient manipulation and beyond,

    S. Zhou, W. Zhang, J. Jiang, W. Zhong, J. Gu, and W. Zhu, “On the convergence of stochastic multi-objective gradient manipulation and beyond,” inAdvances in Neural Information Processing Systems, S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh, Eds., vol. 35,...

  7. [15]

    Direction-oriented multi-objective learning: Simple and provable stochastic algorithms,

    P. Xiao, H. Ban, and K. Ji, “Direction-oriented multi-objective learning: Simple and provable stochastic algorithms,” inProc. Advances Neural Inf. Process. Syst., vol. 36, New Orleans, LA, USA, 2023, pp. 4509– 4533

  8. [16]

    FERERO: A flexible framework for preference-guided multi-objective learning,

    L. Chen, A. F. M. Saif, Y . Shen, and T. Chen, “FERERO: A flexible framework for preference-guided multi-objective learning,” inProc. Advances Neural Inf. Process. Syst., vol. 37, Vancouver, Canada, 2024, pp. 18 758–18 805

  9. [17]

    Joint gradient balancing for data ordering in finite-sum multi-objective optimization,

    H. Yang and J. T. Kwok, “Joint gradient balancing for data ordering in finite-sum multi-objective optimization,” inProc. Int. Conf. Learn. Representations, Singapore, Apr. 2025

  10. [18]

    Variance reduction can improve trade-off in multi-objective learning,

    H. Fernando, L. Chen, S. Lu, P.-Y . Chen, M. Liu, S. Chaudhury, K. Murugesan, G. Liu, M. Wang, and T. Chen, “Variance reduction can improve trade-off in multi-objective learning,” inProc. IEEE Int. Conf. Acoustics, Speech Signal Process., 2024, pp. 6975–6979

  11. [19]

    Proximal gradient methods for multiobjective optimization and their applications,

    H. Tanabe, E. H. Fukuda, and N. Yamashita, “Proximal gradient methods for multiobjective optimization and their applications,”Computational Optimization and Applications, vol. 72, no. 2, pp. 339–361, 2019

  12. [20]

    LibMTL: A python library for deep multi-task learning,

    B. Lin and Y . Zhang, “LibMTL: A python library for deep multi-task learning,”Journal of Machine Learning Research, vol. 24, no. 209, pp. 1–7, 2023

  13. [21]

    Deep hashing network for unsupervised domain adaptation,

    H. Venkateswara, J. Eusebio, S. Chakraborty, and S. Panchanathan, “Deep hashing network for unsupervised domain adaptation,” inPro- ceedings of the IEEE Conference on Computer Vision and Pattern Recognition, jul 2017, pp. 5018–5027

  14. [22]

    R. A. Horn and C. R. Johnson,Matrix Analysis, 2nd ed. Cambridge University Press, 2012. PREPRINT, JULY 2026 8 APPENDIX A summary of notation used in this work is listed in Table III for ease of reference. TABLE III NOTATION AND DESCRIPTIONS. Notation Descriptions x∈R p Model p...

  15. [23]

    Therefore, under the case condition µmin(U ⊤Q⊤ t QtU)< µ t, we have∥q Zt,1 −q Zt,2∥2 <2µ t.Multiplying (93) by1 Rc t , bounding the last term via 1Rc t ∥qZt,1 −q Zt,2∥2 <2µ t, dropping the indicator on the two variance terms, and taking expectations yields E 1Rc t ∥dt −d(x t)∥...

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.