Pith. sign in

REVIEW 2 major objections 5 minor 42 references

The paper claims that domain style, correctly used at the distribution level, is not a nuisance to remove but a navigation signal: an inference-time reweighting of frozen source-specific classifiers by style distance to the target can beat

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 15:24 UTC pith:45OLGCXX

load-bearing objection Worth a serious look: LADDER gives a clean fixed-model recipe for style-guided source reweighting, with honest theory and a real iWildCam gain, but its load-bearing smoothness assumption is only validated on FMoW. the 2 major comments →

arxiv 2607.26458 v1 pith:45OLGCXX submitted 2026-07-29 stat.ML cs.LGmath.STstat.MEstat.TH

Chaos Is a LADDER: Domain Generalization Beyond Invariance via Reweighting

Yuhang Jiang , Fengchuan Zhang , Sanguo Zhang , Guojun Zhu This is my paper
classification stat.ML cs.LGmath.STstat.MEstat.TH MSC 62H3068T05
keywords domain generalizationdomain-dependent prediction rulessource classifier reweightingstyle distributionsoptimal transport distanceunlabeled target covariatesfixed-model inferenceinvariance
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper argues that when the rule connecting causal features to labels varies across domains, domain style should neither be discarded as nuisance nor fed into a pooled classifier as a per-sample shortcut. Instead, style should be summarized at the domain level and used at inference to decide which source-specific prediction rules to trust. The proposed pipeline, LADDER, freezes the learned encoders and source classifiers, builds a style fingerprint from unlabeled target covariates, measures optimal-transport distances to cached source fingerprints, and softmax-weights the K nearest source classifiers. The paper proves finite-sample bounds on the reweighted estimator—parameter error O(eta_src + eta_emp + eta_knn) and excess risk O(R_c^2 E_orc^2)—and reports gains over pooled empirical-risk minimization and invariant prediction: 61.71% test accuracy on a location-grouped wildlife benchmark versus 59.22% for the matched control, and 50.19% versus 49.98% on a satellite benchmark. The significance would be a principled middle path: invariance when rules are stable, distribution-level routing when they shift.

Core claim

Under rule variation, the target domain's style distribution is a navigation variable: the unknown prediction rule beta_* is Lipschitz-continuous in style distribution, so nearby source rules can be weighted by 1-Wasserstein distance. LADDER's reweighted estimator, beta-hat = sum over selected sources of weight_e times beta-hat_e, is within O(eta_src + eta_emp + eta_knn) of the target parameter, where eta_src is finite-sample source estimation error, eta_emp is empirical style-measure error, and eta_knn = (K/E)^(1/d_M) is the interpolation gap to selected sources; excess risk scales as O(R_c^2 E_orc^2). The paper also shows that additive pooled style predictors carry irreducible excess risk

What carries the argument

The central mechanism is the smoothness principle: the distance between two domains' prediction rules is bounded by a Lipschitz constant times the 1-Wasserstein distance between their style distributions. LADDER operationalizes this as distribution-level nearest-neighbor routing: it caches empirical style fingerprints for source domains, forms a target fingerprint from unlabeled covariates, estimates distances with a debiased entropic optimal-transport approximation, selects the K nearest sources, softmax-weights them, and averages the frozen source classifiers. Theorems 2 and 3 decompose the resulting error into source estimation, empirical measure concentration, and KNN interpolation terms

Load-bearing premise

The load-bearing premise is Eq. (1): prediction rules must vary Lipschitz-continuously with style-distribution geometry—if style distance mostly tracks nuisance variation unrelated to rule change, the weights add noise and LADDER degrades toward uniform weighting; the paper also states that no theorem connects its training objective to the oracle representations assumed in the bounds.

What would settle it

On the camera-trap benchmark, compute the same source-geometry diagnostic reported for the satellite data: Spearman correlation between LADDER style distances and bidirectional source-classifier transfer loss over source pairs. Near-zero or negative correlation would falsify the navigation mechanism; the reported gains would then reflect source-classifier diversity rather than style-guided routing. A second check: inject high-variance nuisance coordinates into the style space during training; if LADDER's weights and accuracy do not change, the distance is not using the geometry the theory requ

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • When causal-to-label rules vary and the target lies within source style coverage, unlabeled target covariates alone suffice for adaptation: no target labels, no gradients, and no model-state updates are needed.
  • Additive pooled use of style is structurally insufficient under rule shift, so the failure of style-as-shortcut predictors is not a tuning issue but a model-class limitation.
  • Adding source domains shrinks the KNN interpolation error ((K/E)^(1/d_M)), but only if the smallest source domain remains large enough to keep its style fingerprint reliable.
  • As target summary size and source coverage grow, the parameter error bound vanishes, and excess target risk vanishes quadratically in that bound.
  • The theoretical error separates cleanly into a representation component and an inference-routing component, so improvements in either place translate directly into risk reductions.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The paper validates the smoothness principle on one benchmark only; running the same transferability diagnostic on the camera-trap data would tell whether the larger reported gain comes from style-guided navigation or from source-classifier diversity.
  • Because the inference stage only needs an unlabeled covariate set, LADDER's routing should transfer to any fixed encoder, including foundation-model embeddings, provided representation error stays small in the sense of the paper's perturbation assumption—a testable route to cheaper deployment.
  • The ablation showing that group-level target summaries beat split-level summaries suggests a practical preprocessing rule: partition unlabeled target covariates into coherent environment units before fingerprinting; the optimal granularity is a hyperparameter the paper leaves open.
  • If the smoothness principle fails near or beyond the source support, convex reweighting cannot extrapolate; combining LADDER with an explicit extrapolation mechanism would be a natural next step.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 5 minor

Summary. The paper proposes LADDER, a fixed-model domain-generalization pipeline that learns causal and style encoders, freezes them, fits source-specific classifiers, and at inference uses only an unlabeled target covariate set to estimate a style summary and reweight the frozen source classifiers. The central claim is that when the causal-to-label rule varies across domains, style-distribution geometry can serve as a ladder: domains close in style are assumed to have similar prediction rules (Eq. 1), and reweighting by style distance improves over pooled ERM and invariant prediction. The authors provide an oracle-representation theory (Theorems 1-3) decomposing target parameter error into source-estimation, empirical-Wasserstein, and KNN-interpolation terms, plus a comparison model showing pooled additive predictors and fixed invariant predictors are misspecified when the rule varies. Empirically, LADDER is evaluated on a controlled SCM, FMoW-WILDS, and a location-grouped iWildCam protocol, reporting higher test accuracy and filtered average-group accuracy than protocol-matched routing controls, while acknowledging lower macro-F1 on iWildCam.

Significance. If the smoothness bridge of Eq. (1) holds for the learned representation, LADDER is a conceptually meaningful contribution: it moves beyond invariant prediction by using domain-level style summaries to navigate among domain-dependent rules, with no target labels and no model updates. The paper includes explicit assumptions, full proofs in the supplement, protocol-matched controls (UniformExperts, NearestExpert, SampleGate), and an honest FMoW diagnostic linking style distance to source-classifier transferability. The error decomposition in Theorem 2 is clean and the comparison model in Corollaries 1-2 is informative. The practical significance, however, depends on two currently open points: the theoretical guarantees are proved for oracle representations and are not connected to the actual Eq. (2) training objective, and the empirical mechanism is validated only on FMoW, not on iWildCam where the headline gain appears.

major comments (2)
  1. [3.2, Assumption 5; 2.2, Eq. (2)] Theorems 2-3 are proved for oracle representations satisfying Assumptions 1-3; Corollary 3 is conditional on Assumption 5. The paper explicitly concedes that Assumption 5 is a modular robustness interface, not a theorem that a particular training objective recovers the oracle representation. No result connects the Eq. (2) objective to Assumptions 1-5, so the guarantees do not currently apply to the model actually trained. This is load-bearing: if Eq. (2) produces representation error outside the sub-Gaussian perturbation model, the additive term in Corollary 3 does not control it. Please either (a) prove conditions under which a tractable variant of Eq. (2) yields representations satisfying Assumption 5 or Assumptions 1-2, or (b) reframe the theorems explicitly as conditional guarantees for any representation satisfying the interface and add an empirical check that these conditions hold
  2. [4.2, Table 2; Supplement Fig. 9] The load-bearing mechanism is the smoothness bridge of Eq. (1)/Assumption 2. The only direct empirical validation is the FMoW diagnostic in Supplement Fig. 9 (Spearman rho = 0.70 across 55 non-independent source pairs), which the paper itself calls descriptive. No analogous diagnostic is reported for iWildCam, where the strongest gain appears (Table 2: LADDER-group 61.71 vs. SampleGate-group 59.22 test accuracy; 50.01 vs. 48.29 filtered avg). Since LADDER-split collapses to 41.19 and grouping alone improves SampleGate over SampleGate (59.22 vs. 58.00), the iWildCam gain could in principle be driven by the location-level grouping protocol or by K/tau choices rather than by distance-based reweighting. I request an iWildCam version of Supplement Fig. 9, e.g., style distance vs. bidirectional source-classifier transfer loss across camera locations, or an equivalent test. Without it, the empi
minor comments (5)
  1. [3.1, Theorems 2-3] Notation overload: E denotes both the number of source domains and the expectation in E norm. This makes statements like E to infinity and E of norm confusing. Use a distinct symbol for expectation throughout.
  2. [2.3 and Supplement A.3] The theory analyzes exact W1 distances, but real-data distances use a balanced debiased Sinkhorn approximation; the supplement states that entropic and solver errors are not included in the rates. Please state this gap explicitly in the main text near Eq. (4) and, if possible, report an empirical estimate of the approximation error.
  3. [4.2, Table 2] LADDER-group's macro F1 (11.61) is lower than SampleGate (14.40) and SampleGate-group (12.00). The paper acknowledges this, but the discussion should address whether reweighting systematically sacrifices rare classes on a severely long-tailed benchmark, since overall accuracy alone can overstate the practical value for iWildCam.
  4. [Supplement Fig. 9] The Spearman rho values are descriptive and based on dependent domain pairs; the supplement notes this. The main text should keep the same caution and avoid wording that implies a causal validation of Eq. (1) from these correlations alone.
  5. [4.2, Supplement Fig. 12] On iWildCam, K=5 gives higher test accuracy and macro F1 than the main K=6 configuration, while K=6 maximizes filtered average-group accuracy. Since the headline metrics move in opposite directions with K, please state explicitly how K=6 was fixed (e.g., a validation-set rule or a pre-specified robustness protocol) and clarify that the reported KNN sweep was not used to select K on the test set.

Circularity Check

0 steps flagged

No significant circularity: explicit assumptions, conditional bounds, no fitted quantity relabeled as prediction.

full rationale

The paper's derivation chain is self-contained conditional on explicitly stated assumptions. The load-bearing smoothness principle, Eq. (1), is introduced as an assumption (Assumption 2 in Section 3.1: "There is an L_beta-Lipschitz map Psi* from domain-level style distributions to GLM parameters"), not derived from the method or from data. Theorems 2 and 3 then propagate this assumption through standard ingredients: GLM strong convexity, Fournier–Guillin empirical Wasserstein rates, and a KNN order-statistic argument. The quantities eta_src, eta_emp, and eta_knn are theoretical convergence rates, not fitted parameters, and the inference weights w_e are computed from style distances without using target labels or transfer-loss values. No fitted parameter is renamed as a prediction. Theorem 1's conclusion M = P D follows from a joint diagonalization condition and is an identifiability lemma with explicit assumptions. The admitted gap that Assumption 5 is "a modular robustness interface rather than a theorem that a particular training objective recovers the oracle representation" (Section 3.2) is a limitation in connecting the training objective to the oracle theory, not a circular equivalence. The paper makes no load-bearing self-citation: the cited theoretical rates are external (Fournier and Guillin 2015) and the WILDS benchmark results are externally defined. The FMoW source-geometry diagnostic (Spearman rho=0.70, Supplement Fig. 9) is descriptive support for Assumption 2, and the absence of an equivalent iWildCam diagnostic is an empirical validation gap, not a definitional or fitted-input circularity. Overall, the results depend on an unproven but honestly stated assumption; that is a correctness/validation risk, not circular reasoning.

Axiom & Free-Parameter Ledger

6 free parameters · 7 axioms · 1 invented entities

The pipeline rests on a chain of assumed structures: a causal/style decomposition that the training objective is not proven to deliver (authors admit Assumption 5 is an interface, not a recovery theorem); a Lipschitz map from style distributions to rules (Eq. 1/Assumption 2); GLM strong convexity and bounded features (Assumption 3); and a dM-dimensional Ahlfors-regular domain manifold for the (K/E)^{1/dM} term. Free parameters are the per-dataset routing hyperparameters (K, τ), the five loss weights in Eq. (2), Sinkhorn blur/scaling/m, target-summary granularity, and source-filtering thresholds. These are fixed per dataset rather than learned, but the K values coincide with the best-performing ablation points, and the location-level grouping choice on iWildCam drives the largest reported gain.

free parameters (6)
  • K (neighbor count) = 4 (simulation), 5 (FMoW), 6 (iWildCam)
    Controls how many source classifiers are retained at inference. Per-dataset protocol values align with the best-performing points in the reported ablations: FMoW K=5 'gives the strongest mean performance' and iWildCam K=6 'gives the strongest filtered average-group accuracy.'
  • τ (softmax temperature) = 0.1 (sim), 0.5 (FMoW), 1.0 (iWildCam)
    Sharpness of the weight vector over source distances; chosen per dataset by 'neighborhood-size and target-summary reliability considerations' and the simulation ablation shows τ strongly controls how much navigation occurs.
  • Loss weights (λ_s, λ_a, λ_o, λ_r) = (0.15, 0.03, 5e-4, 1e-5)
    Hand-chosen weights in Eq. (2) balancing style-domain classification, adversarial suppression (with the disclosed λ_a^2 gradient scale), orthogonality, and regularization; no selection procedure is reported.
  • Sinkhorn parameters (blur, scaling, m) = blur=0.05, scaling=0.8, m=1024
    Distance-estimator hyperparameters for the debiased Sinkhorn divergence used on real data; the theory covers exact W1 only, so these introduce unquantified approximation error.
  • Target-summary granularity = split-level (FMoW, iWildCam-split); location-level (iWildCam-group)
    Material protocol decision: iWildCam-group reaches 61.71% test accuracy vs 41.19% for LADDER-split, so the headline real-data result depends on this grouping choice.
  • Source-domain filtering thresholds = min size 512; 11 FMoW years; 64 iWildCam locations
    Excludes small source domains before training classifiers and caching fingerprints; changes the source pool over which the ensemble operates.
axioms (7)
  • domain assumption Smoothness principle (Eq. 1): d_F(η_e, η_e') ≤ L_η W1(ν_e, ν_e') — prediction rules are Lipschitz in style-distribution distance.
    The load-bearing premise: if style distance does not track rule similarity, the reweighting signal is noise. The only empirical check is a descriptive FMoW diagnostic (ρ = 0.70, Supplement Fig. 9); no iWildCam validation is given.
  • domain assumption Assumption 1 (oracle latent structure): X = A_c Z_c + A_s Z_s + ε with domain-invariant causal/noise covariance, diagonal Λ_e, jointly separating variance profiles.
    Needed for Theorem 1 identifiability; no result shows the LADDER training objective (Eq. 2) yields extractors that diagonalize the domain-varying style covariance.
  • domain assumption Assumption 2 (Ahlfors-regular dM-dimensional meta-distribution; L_β-Lipschitz Ψ*; i.i.d. source style laws).
    Required for the coverage term (K/E)^{1/dM} and for W1 concentration; the manifold dimension dM is unobserved and not estimated.
  • standard math Assumption 3 (GLM regularity): bounded causal features, uniform μ-strong convexity of population risks, L_H-Lipschitz Hessians.
    Standard M-estimation regularity used in Lemmas 1–3 and Theorem 3; plausible for linear GLMs, unverified for the empirical encoder outputs.
  • domain assumption Assumption 4 (comparison shift model): scalar latent coordinate c with Var(c) > 0, Z_c ⊥ (c, Z_s), c_inv ≠ c_0.
    Diagnostic model for Corollaries 1–2; the authors explicitly call it 'intentionally narrower' and separate it from the main LADDER analysis.
  • domain assumption Assumption 5 (sub-Gaussian representation error; perturbed empirical risk μ_rep-strongly convex).
    Modular interface; the authors concede it is 'not a theorem that a particular training objective recovers the oracle representation,' leaving the theory-reality gap open.
  • domain assumption Structural causal model in §2.1: X = G_e(Z_c, Z_s, ε_x), Y ∼ η_e(·|Z_c).
    Formalizes the rule-varying setting; the causal/style decomposition is operational disentanglement, explicitly not statistical independence.
invented entities (1)
  • Domain coordinate c (comparison model) no independent evidence
    purpose: Scalar latent domain position used to instantiate the pooled-joint lower bound and the invariant-prediction comparison (Corollaries 1–2).
    A diagnostic modeling device, not claimed to exist in real data; no falsifiable prediction is attached. The paper introduces no new physical entities, particles, dimensions, or forces beyond this analytical construction.

pith-pipeline@v1.3.0-daily-deepseek · 5661 in / 6955 out tokens · 269809 ms · 2026-08-01T15:24:14.504739+00:00 · methodology

0 comments
read the original abstract

Domain generalization (DG) aims to learn from multiple source domains and generalize to unseen target domains. Most DG methods pursue invariance: they seek a causal representation whose prediction rule is invariant across domains. This principle is effective when the causal mechanism is stable, but becomes restrictive when the domain itself modulates how causal content maps to the response. In this case, directly feeding domain style into the predictor can create misleading shortcuts, since style does not by itself cause the response. Yet the apparent chaos of multiple styles can become a ladder: style can locate the unseen target domain among source domains and guide which domain-dependent prediction rules should be trusted. We propose \emph{Latent Adaptive Domain Disentanglement and Environment Reweighting} (LADDER), a fixed-model DG pipeline that learns causal/style representations, freezes the encoders, fits source-specific classifiers, and uses an unlabeled target-domain covariate set only at inference to compute weights over these fixed classifiers, with no target labels or model-state updates. We establish theoretical guarantees for source reweighting and validate LADDER on simulations, FMoW, and a location-grouped iWildCam protocol, with gains in overall and group-averaged accuracy.

Figures

Figures reproduced from arXiv: 2607.26458 by Fengchuan Zhang, Guojun Zhu, Sanguo Zhang, Yuhang Jiang.

Figure 1
Figure 1. Figure 1: General structural causal model. The edge [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of LADDER. Source domains are encoded into causal representations for source-specific classifiers and style representations for source-domain fingerprints. At prediction, the fixed style encoder maps unlabeled target covariates to a style summary, compares it with stored source fingerprints, and changes only weights over fixed source classifiers. 2.2 Model Training: Encoders and Classifiers Model … view at source ↗
Figure 3
Figure 3. Figure 3: FMoW-WILDS accuracy and worst-region accuracy on the ID validation, OOD validation, and OOD test splits. Error [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: FMoW-WILDS accuracy/average-region trade-off, OOD degradation, and representative [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: FMoW-WILDS representation mechanism and source-geometry diagnostic. Panels (a,b) visualize with t-SNE [van der [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: iWildCam-WILDS test accuracy and filtered average [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: iWildCam-WILDS trade-off and LADDER target￾summary ablation. LADDER-group lies in the high-accuracy, high-robustness region, and the split-to-group ablation shows that camera-location-level summaries matter for this benchmark. All methods use ImageNet-pretrained ResNet-50 [He et al. 2016, Deng et al. 2009], 64 source locations, class-balanced cross-entropy with parameter 0.999 [Cui et al. 2019], and five s… view at source ↗
Figure 1
Figure 1. Figure 1: Simulation results over 100 independent repetitions. [PITH_FULL_IMAGE:figures/full_fig_p013_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Synthetic simulation KNN source-selection ablation over [PITH_FULL_IMAGE:figures/full_fig_p014_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Synthetic simulation ablation for target-summary sample size and the temperature [PITH_FULL_IMAGE:figures/full_fig_p015_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Supplementary style-conditioned pooled ablation in the controlled SCM. Stronger [PITH_FULL_IMAGE:figures/full_fig_p016_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: FMoW-WILDS accuracy and worst-region accuracy on the ID validation, OOD [PITH_FULL_IMAGE:figures/full_fig_p017_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: FMoW-WILDS accuracy/average-region trade-off, OOD degradation, and repre [PITH_FULL_IMAGE:figures/full_fig_p018_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: FMoW-WILDS KNN source-selection ablation over five random seeds. The main [PITH_FULL_IMAGE:figures/full_fig_p018_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Full FMoW-WILDS t-SNE visualization. Panels (a,b) show the causal representa [PITH_FULL_IMAGE:figures/full_fig_p019_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: FMoW-WILDS source-geometry diagnostic over all 55 source-domain pairs. The [PITH_FULL_IMAGE:figures/full_fig_p020_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: iWildCam-WILDS test accuracy and filtered average-group accuracy. [PITH_FULL_IMAGE:figures/full_fig_p021_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: iWildCam-WILDS accuracy/robustness trade-off and [PITH_FULL_IMAGE:figures/full_fig_p021_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: iWildCam-WILDS KNN source-selection ablation over three random seeds. The [PITH_FULL_IMAGE:figures/full_fig_p022_12.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

42 extracted references · 7 linked inside Pith

  1. [1]

    Ahuja, K.; Caballero, E.; Zhang, D.; Gagnon-Audet, J.-C.; Bengio, Y.; Mitliagkas, I.; and Rish, I. 2021. Invariance Principle Meets Information Bottleneck for Out-of-Distribution Generalization. arXiv:2106.06607

    Pith/arXiv arXiv 2021

  2. [2]

    A.; Fischer, I.; Dillon, J

    Alemi, A. A.; Fischer, I.; Dillon, J. V.; and Murphy, K. 2017. Deep Variational Information Bottleneck. In International Conference on Learning Representations

    2017

  3. [3]

    Arjovsky, M.; Bottou, L.; Gulrajani, I.; and Lopez-Paz, D. 2019. Invariant Risk Minimization. arXiv:1907.02893

    Pith/arXiv arXiv 2019

  4. [4]

    Arpit, D.; Wang, H.; Zhou, Y.; and Xiong, C. 2022. Ensemble of Averages: Improving Model Selection and Boosting Performance in Domain Generalization. In Advances in Neural Information Processing Systems

    2022

  5. [5]

    Beery, S.; Cole, E.; Parker, J.; Perona, P.; and Winner, K. 2020. The iWildCam 2020 Competition Dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops

    2020

  6. [6]

    Blanchard, G.; Lee, G.; and Scott, C. 2011. Generalizing from Several Related Classification Tasks to a New Unlabeled Sample. In Advances in Neural Information Processing Systems

    2011

  7. [7]

    Christie, G.; Fendley, N.; Wilson, J.; and Mukherjee, R. 2018. Functional Map of the World. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition

    2018

  8. [8]

    Cui, Y.; Jia, M.; Lin, T.-Y.; Song, Y.; and Belongie, S. 2019. Class-Balanced Loss Based on Effective Number of Samples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    2019

  9. [9]

    Deng, J.; Dong, W.; Socher, R.; Li, L.-J.; Li, K.; and Fei-Fei, L. 2009. ImageNet: A Large-Scale Hierarchical Image Database. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition

    2009

  10. [10]

    u gelgen, J.; Hassani, H.; Singh, S.; Pappas, G. J.; and Sch \

    Eastwood, C.; Robey, A.; von K \"u gelgen, J.; Hassani, H.; Singh, S.; Pappas, G. J.; and Sch \"o lkopf, B. 2022. Probable Domain Generalization via Quantile Risk Minimization. In Advances in Neural Information Processing Systems

    2022

  11. [11]

    L.; Maritsch, M.; Okanovic, P.; Thaeter, G.; Buhmann, J

    Foll, S.; Dubatovka, A.; Ernst, E.; Chau, S. L.; Maritsch, M.; Okanovic, P.; Thaeter, G.; Buhmann, J. M.; Wortmann, F.; and Muandet, K. 2023. Gated Domain Units for Multi-source Domain Generalization. Transactions on Machine Learning Research

    2023

  12. [12]

    Fournier, N.; and Guillin, A. 2015. On the Rate of Convergence in Wasserstein Distance of the Empirical Measure. Probability Theory and Related Fields, 162(3--4): 707--738

    2015

  13. [13]

    Gulrajani, I.; and Lopez-Paz, D. 2021. In Search of Lost Domain Generalization. In International Conference on Learning Representations

    2021

  14. [14]

    He, K.; Zhang, X.; Ren, S.; and Sun, J. 2016. Deep Residual Learning for Image Recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition

    2016

  15. [15]

    W.; Sagawa, S.; Marklund, H.; Xie, S

    Koh, P. W.; Sagawa, S.; Marklund, H.; Xie, S. M.; Zhang, M.; Balsubramani, A.; Hu, W.; Yasunaga, M.; Phillips, R. L.; Gao, I.; Lee, T.; David, E.; Stavness, I.; Guo, W.; Earnshaw, B.; Haque, I. S.; Beery, S.; Leskovec, J.; Kundaje, A.; Pierson, E.; Levine, S.; Finn, C.; and Liang, P. 2021. WILDS: A Benchmark of in-the-Wild Distribution Shifts. In Internat...

    2021

  16. [16]

    L.; and Courville, A

    Krueger, D.; Caballero, E.; Jacobsen, J.-H.; Zhang, A.; Binas, J.; Priol, R. L.; and Courville, A. 2021. Out-of-Distribution Generalization via Risk Extrapolation (REx). In Proceedings of the 38th International Conference on Machine Learning, 5815--5826

    2021

  17. [17]

    Li, S.; and Zhang, L. 2023. Multi-dimensional Domain Generalization with Low-rank Structures. arXiv:2309.09555

    Pith/arXiv arXiv 2023

  18. [18]

    Mancini, M.; Rota Bulo, S.; Caputo, B.; and Ricci, E. 2018. Best Sources Forward: Domain Generalization Through Source-Specific Nets. In 2018 25th IEEE International Conference on Image Processing (ICIP), 1353--1357. IEEE

    2018

  19. [19]

    Nguyen, T.; Do, K.; Duong, B.; and Nguyen, T. 2024. Domain Generalisation via Risk Distribution Matching. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision

    2024

  20. [20]

    Peters, J.; B \"u hlmann, P.; and Meinshausen, N. 2016. Causal Inference by Using Invariant Prediction: Identification and Confidence Intervals. Journal of the Royal Statistical Society: Series B, 78(5): 947--1012

    2016

  21. [21]

    Rame, A.; Dancette, C.; and Cord, M. 2022. Fishr: Invariant Gradient Variances for Out-of-Distribution Generalization. In Proceedings of the 39th International Conference on Machine Learning

    2022

  22. [22]

    Rosenfeld, E.; Ravikumar, P.; and Risteski, A. 2021. The Risks of Invariant Risk Minimization. In International Conference on Learning Representations

    2021

  23. [23]

    a usler, D.; B \

    Rothenh \"a usler, D.; B \"u hlmann, P.; Meinshausen, N.; and Peters, J. 2021. Anchor Regression: Heterogeneous Data Meet Causality. Journal of the Royal Statistical Society: Series B, 83(2): 215--246

    2021

  24. [24]

    Shazeer, N.; Mirhoseini, A.; Maziarz, K.; Davis, A.; Le, Q.; Hinton, G.; and Dean, J. 2017. Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer. arXiv preprint arXiv:1701.06538

    Pith/arXiv arXiv 2017

  25. [25]

    van der Maaten, L.; and Hinton, G. 2008. Visualizing Data using t-SNE. Journal of Machine Learning Research, 9: 2579--2605

    2008

  26. [26]

    Xiao, Z.; and Snoek, C. G. M. 2024. Beyond Model Adaptation at Test Time: A Survey. arXiv:2411.03687

    Pith/arXiv arXiv 2024

  27. [27]

    Zhang, H.; Zhang, Y.-F.; Liu, W.; Weller, A.; Sch \"o lkopf, B.; and Xing, E. P. 2022. Towards Principled Disentanglement for Domain Generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    2022

  28. [28]

    Zhang, Y. 2021. A Survey of Unsupervised Domain Adaptation for Visual Recognition. arXiv:2112.06745

    Pith/arXiv arXiv 2021

  29. [29]

    Zheng, S.; Lin, Y.; Zhou, Y.; and Huang, J. 2026. Domain-Specific Nonparametric Regression for Domain Generalization. Journal of the American Statistical Association

    2026

  30. [30]

    Zhou, K.; Liu, Z.; Qiao, Y.; Xiang, T.; and Loy, C. C. 2022. Generalizing to Unseen Domains: A Survey on Domain Generalization. IEEE Transactions on Pattern Analysis and Machine Intelligence

    2022

  31. [31]

    Zhou, K.; Yang, Y.; Qiao, Y.; and Xiang, T. 2021. Domain Adaptive Ensemble Learning. IEEE Transactions on Image Processing, 30: 8008--8018

    2021

  32. [32]

    Zhu, Y.; Deng, N.; Shi, N.; Gangrade, A.; and Scott, C. 2025. Domain Generalization: A Tale of Two ERMs. arXiv:2510.04441

    arXiv 2025

  33. [33]

    Zuo, S.; and Wang, Y. 2026. Environment-Adaptive Covariate Selection: Learning When to Use Spurious Correlations for Out-of-Distribution Prediction. arXiv:2601.02322

    Pith/arXiv arXiv 2026

  34. [34]

    Biau, G.; and Devroye, L. 2015. Lectures on the Nearest Neighbor Method. Springer

    2015

  35. [35]

    Boucheron, S.; Lugosi, G.; and Massart, P. 2013. Concentration Inequalities: A Nonasymptotic Theory of Independence. Oxford University Press

    2013

  36. [36]

    Devroye, L.; Gy \"o rfi, L.; and Lugosi, G. 1996. A Probabilistic Theory of Pattern Recognition. Springer

    1996

  37. [37]

    McCullagh, P.; and Nelder, J. A. 1989. Generalized Linear Models. Chapman and Hall, 2 edition

    1989

  38. [38]

    Tropp, J. A. 2012. User-Friendly Tail Bounds for Sums of Random Matrices. Foundations of Computational Mathematics, 12(4): 389--434

    2012

  39. [39]

    van der Vaart, A. W. 1998. Asymptotic Statistics. Cambridge University Press

    1998

  40. [40]

    Vershynin, R. 2018. High-Dimensional Probability: An Introduction with Applications in Data Science. Cambridge University Press

    2018

  41. [41]

    Villani, C. 2009. Optimal Transport: Old and New. Springer

    2009

  42. [42]

    Wainwright, M. J. 2019. High-Dimensional Statistics: A Non-Asymptotic Viewpoint. Cambridge University Press

    2019