pith. machine review for the scientific record. sign in

arxiv: 2605.11394 · v1 · submitted 2026-05-12 · 📊 stat.ML · cs.AI· cs.LG· stat.AP· stat.ME

Recognition: 1 theorem link

· Lean Theorem

Spatial Adapter: Structured Spatial Decomposition and Closed-Form Covariance for Frozen Predictors

Hao-Yun Huang, Wei-Ying Wu, Wen-Ting Wang, Xuan-Chun Wang

Pith reviewed 2026-05-13 01:35 UTC · model grok-4.3

classification 📊 stat.ML cs.AIcs.LGstat.APstat.ME
keywords spatial adapterresidual decompositionfrozen predictorsspatial covariancekrigingorthonormal basisADMM optimizationparameter-efficient adaptation
0
0 comments X

The pith

The Spatial Adapter equips any frozen predictor with a structured spatial representation of its residual field and a closed-form spatial covariance.

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

The paper claims that a lightweight post-hoc layer can extract usable spatial structure from the residuals of any already-trained predictor, whether linear or deep, without touching the original parameters. It does this by learning a spatially regularized orthonormal basis together with per-sample scores through a mini-batch optimization that enforces smoothness, sparsity, and orthogonality. The resulting low-rank-plus-noise model yields an explicit covariance for the residuals, which in turn supports spatial interpolation at new locations and uncertainty estimates. A sympathetic reader would care because the approach promises to add spatial awareness to large deployed models at far lower cost than retraining or fine-tuning them from scratch.

Core claim

The Spatial Adapter operates as a cascade second stage on the residuals of a frozen first-stage predictor. It jointly learns a spatially regularized orthonormal basis and per-sample scores via a tractable mini-batch ADMM procedure. Smoothness, sparsity, and orthogonality constraints convert a generic low-rank factorization into an identifiable spatial representation whose induced residual covariance admits a closed-form low-rank-plus-noise estimator. The effective rank is selected data-adaptively by spectral thresholding while the nominal rank K serves only as an optimization upper bound. This covariance enables kriging-style spatial prediction at unobserved locations together with plug-in 0

What carries the argument

The spatially regularized orthonormal basis and per-sample scores, jointly optimized via mini-batch ADMM on residuals, which directly induce the closed-form low-rank-plus-noise spatial covariance.

If this is right

  • Kriging-style prediction becomes available at unobserved locations using the closed-form covariance.
  • Plug-in uncertainty quantification follows directly from the same covariance estimator.
  • The method applies unchanged to frozen first stages ranging from linear models to deep spatiotemporal and vision backbones.
  • The added representation remains compact, using fewer than K(N+T) parameters plus a small residual-trend network.

Where Pith is reading between the lines

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

  • The same residual-decomposition step could be inserted after any predictor whose outputs exhibit unmodeled spatial autocorrelation.
  • The closed-form covariance might serve as a drop-in component inside larger Gaussian-process or hybrid spatial models.
  • Spectral thresholding for rank selection could be tested for stability across repeated random splits of the same spatial dataset.

Load-bearing premise

The residuals of the frozen first-stage predictor contain identifiable spatial structure that a joint optimization of a spatially regularized orthonormal basis and per-sample scores can recover.

What would settle it

If adding the adapter produces no measurable gain in spatial-holdout prediction accuracy on Weather2K relative to the frozen baseline alone, or if the induced covariance matrix deviates substantially from the empirical covariance of the observed residuals.

Figures

Figures reproduced from arXiv: 2605.11394 by Hao-Yun Huang, Wei-Ying Wu, Wen-Ting Wang, Xuan-Chun Wang.

Figure 1
Figure 1. Figure 1: Spatial Adapter overview. The adapter extracts an explicit low-rank spatial representation Φb(s) ⊤αbj from the residual field of a frozen first-stage predictor, supporting prediction together with closed-form covariance estimation and downstream uncertainty quantification. Stage-2 ADMM details: Algorithm 1. 2. Structured covariance induced by the representation. Φb directly induces a closed-form low-rank-p… view at source ↗
Figure 2
Figure 2. Figure 2: Synthetic mechanism validation (rank-1 DGP, representative seed). Left: 3 [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Held-out Weather2K kriging at one test time. Left: STDK extrapolation baseline. Middle: adapter kriging surface — STDK plus the learned low-rank correction Φb(s) ⊤αbj , interpolated to held-out locations via thin-plate splines. Right: no-holdout reference — the same pipeline refit with the 20% held-out stations included in training, so the basis is read off in-sample (no TPS interpolation). All three panel… view at source ↗
Figure 4
Figure 4. Figure 4: Per-station held-out RMSE on Weather2K. Left: RMSE ratio (adapter / STDK) overlaid on the held-out-station map; cool colours indicate the adapter outperforms the backbone. Right: paired RMSE trajectories, stations sorted by STDK RMSE — the gap between the two curves shows consistent per-station improvement across the 20% held-out stations, not an average driven by a few outliers. 4.2 Wheat Head: Basis Tran… view at source ↗
Figure 5
Figure 5. Figure 5: GWHD within-image refinement (ResNet-152, shared scale [0, 1]): (a) image + GT boxes, (b) GT patch mask, (c) backbone probability, (d) adapter-refined (spatially sharpened, aligns with (b)). the adapter projects each test image’s observed patch labels onto the train-learned basis Φb, so reported metrics measure how well Φb spans the test-time residual structure rather than predictive classification (a labe… view at source ↗
Figure 6
Figure 6. Figure 6: ADMM convergence on the synthetic benchmark (seed 42). [PITH_FULL_IMAGE:figures/full_fig_p020_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Per-experiment cumulative-variance curves for the rank-selection rule of Section 2.3. Vertical lines mark the chosen K at τvar = 0.9. 22 [PITH_FULL_IMAGE:figures/full_fig_p022_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Station split (observed vs. held-out). N Weather2K data-split ablation Complementary to the held-out spatial prediction in Section 4.1, this appendix reports a time￾split ablation with all stations observed. We run two contrasting train/val/test allocations, 80/10/10 and 10/10/80, to examine how the split affects both prediction accuracy and spatial covariance recovery: varying the training proportion reve… view at source ↗
Figure 9
Figure 9. Figure 9: Signed low-rank correction Φb(s) ⊤αbj at the same test time as [PITH_FULL_IMAGE:figures/full_fig_p025_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Estimated spatial covariance Σbr along the regularization path on the synthetic benchmark of Appendix I (median-representative seed, shared colour scale). Three regimes are visible: under-regularization injects spurious off-diagonal structure; λ ⋆ recovers the ground-truth rank-1 covariance; over-regularization collapses toward the noise floor. The accompanying basis￾alignment / CovFrob path summary is in… view at source ↗
read the original abstract

We present the Spatial Adapter, a parameter-efficient post-hoc layer that equips any frozen first-stage predictor with a structured spatial representation of its residual field and an induced closed-form spatial covariance. The adapter operates as a cascade second stage on residuals, jointly learning a spatially regularized orthonormal basis and per-sample scores via a tractable mini-batch ADMM procedure, without modifying any first-stage parameter. Because the first-stage parameters are frozen, the adapter does not retrain the backbone; its role is to supply a compressed distributional summary of the residual field. Smoothness, sparsity, and orthogonality together turn a generic low-rank factorization into an identifiable spatial representation whose induced residual covariance admits a closed-form low-rank-plus-noise estimator; the effective rank is determined data-adaptively by spectral thresholding, while the nominal rank K is an optimization-side upper bound only. This covariance enables kriging-style spatial prediction at unobserved locations, with plug-in uncertainty quantification as a secondary downstream use. Across synthetic data, Weather2K for spatial-holdout prediction, and GWHD patch grids as a basis-transferability diagnostic, the adapter recovers residual spatial structure when paired with frozen first stages from linear models to deep spatiotemporal and vision backbones; the added representation uses fewer than K(N+T) parameters alongside a compact residual-trend network.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 1 minor

Summary. The manuscript introduces the Spatial Adapter, a parameter-efficient post-hoc layer for any frozen first-stage predictor. It jointly learns a spatially regularized orthonormal basis and per-sample scores for the residual field via mini-batch ADMM, derives an induced closed-form low-rank-plus-noise covariance estimator, and selects effective rank data-adaptively by spectral thresholding. The nominal rank K serves only as an upper bound. Experiments on synthetic data, Weather2K spatial-holdout prediction, and GWHD patch grids show the adapter recovers residual spatial structure across linear models to deep spatiotemporal and vision backbones while using fewer than K(N+T) parameters plus a compact residual-trend network, enabling kriging-style prediction and uncertainty quantification without retraining the backbone.

Significance. If the identifiability of the basis, the validity of the closed-form covariance, and the reliability of spectral thresholding hold under the stated assumptions, the adapter would offer a practical, low-parameter way to equip pretrained models with explicit spatial covariance modeling. This could be useful in spatiotemporal forecasting and vision tasks where frozen backbones are standard and residuals exhibit recoverable spatial correlation, providing efficiency advantages over full fine-tuning or separate spatial models.

major comments (3)
  1. [Abstract and identifiability discussion] Abstract and § on identifiability: the central claim that smoothness, sparsity, and orthogonality yield an 'identifiable spatial representation' is load-bearing but unsupported by any theorem, sufficient conditions (e.g., minimum SNR, degree of spatial smoothness, or separation from non-spatial components), or analysis showing that mini-batch ADMM recovers the true basis rather than noise artifacts for arbitrary residual fields.
  2. [Covariance estimation and rank selection] Covariance section: although the closed-form expression given the basis is non-circular, the data-adaptive spectral thresholding for effective rank is fitted to the same residuals; the manuscript does not analyze bias, consistency, or when thresholding recovers signal rank versus noise, which directly affects the downstream covariance estimator and kriging claims.
  3. [Experiments] Experimental validation: the synthetic experiments do not probe failure regimes (unstructured noise, low SNR, or misspecified first-stage residuals) where the assumption that residuals contain recoverable spatial structure may break; this leaves the general claim that the adapter works for arbitrary frozen predictors untested at its weakest point.
minor comments (1)
  1. [Abstract] Notation for N and T in the parameter count claim is not defined on first use in the abstract, reducing clarity for readers unfamiliar with the spatial grid dimensions.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed comments. We address each major point below with clarifications and proposed revisions. Where the manuscript lacks formal analysis, we will add discussion of limitations and supporting experiments.

read point-by-point responses

  1. Referee: [Abstract and identifiability discussion] Abstract and § on identifiability: the central claim that smoothness, sparsity, and orthogonality yield an 'identifiable spatial representation' is load-bearing but unsupported by any theorem, sufficient conditions (e.g., minimum SNR, degree of spatial smoothness, or separation from non-spatial components), or analysis showing that mini-batch ADMM recovers the true basis rather than noise artifacts for arbitrary residual fields.

    Authors: We agree that a formal identifiability theorem with explicit sufficient conditions (e.g., minimum SNR or separation) is absent. The manuscript motivates the claim heuristically: the joint enforcement of spatial smoothness (via the regularizer), sparsity, and orthonormality on the basis distinguishes it from generic low-rank factorizations, reducing the chance of recovering unstructured noise. The mini-batch ADMM optimizes this constrained objective directly on residuals. In revision we will expand the relevant section with a heuristic derivation of uniqueness up to sign, references to related identifiability results in regularized matrix factorization, and an explicit limitations paragraph noting regimes where recovery may fail (low SNR, non-spatial residuals). No new theorem will be added, as deriving one under general residual fields exceeds the current scope. revision: partial

  2. Referee: [Covariance estimation and rank selection] Covariance section: although the closed-form expression given the basis is non-circular, the data-adaptive spectral thresholding for effective rank is fitted to the same residuals; the manuscript does not analyze bias, consistency, or when thresholding recovers signal rank versus noise, which directly affects the downstream covariance estimator and kriging claims.

    Authors: The spectral thresholding is a practical, data-driven heuristic applied to the residual Gram matrix after basis estimation. We will add a dedicated subsection deriving the bias of the resulting low-rank-plus-noise covariance estimator under the model assumptions, showing consistency of the thresholded rank when the gap between signal and noise eigenvalues exceeds a constant factor (with a brief proof sketch). We will also report the effective rank selected across experiments and discuss sensitivity to the threshold parameter. This addresses the impact on kriging variance estimates. revision: partial

  3. Referee: [Experiments] Experimental validation: the synthetic experiments do not probe failure regimes (unstructured noise, low SNR, or misspecified first-stage residuals) where the assumption that residuals contain recoverable spatial structure may break; this leaves the general claim that the adapter works for arbitrary frozen predictors untested at its weakest point.

    Authors: We accept that the current synthetic suite does not systematically include unstructured noise or very low-SNR regimes. In the revision we will add two new synthetic panels: (i) residuals drawn from isotropic Gaussian noise with increasing variance, and (ii) first-stage predictors deliberately misspecified by omitting key spatial covariates. These will quantify when the adapter's spatial basis collapses to noise and when kriging performance degrades to the baseline. The existing experiments already span linear to deep backbones and varying spatial correlation strengths, but the new cases will better bound the method's applicability. revision: yes

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper defines the Spatial Adapter as a fitting procedure that jointly optimizes a spatially regularized orthonormal basis and per-sample scores on residuals via mini-batch ADMM, then supplies a mathematically derived closed-form low-rank-plus-noise covariance estimator from those fitted quantities. The closed-form covariance is a direct algebraic consequence of the basis and scores (non-circular given the fit), while data-adaptive spectral thresholding for rank is an explicit algorithmic component of the method rather than a renamed prediction. Experiments on synthetic data, Weather2K, and GWHD provide external empirical checks against held-out spatial structure, so the central claims do not reduce to self-definition, tautological renaming, or load-bearing self-citation. The derivation remains self-contained as a proposed estimator with stated assumptions about residual structure.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

The central claim rests on the existence of recoverable spatial structure in residuals, the validity of the ADMM optimization for joint basis and score learning, and the correctness of the induced covariance formula; these are not independently verified in the abstract.

free parameters (2)
  • nominal rank K
    Optimization-side upper bound on the number of basis functions; chosen by the user and affects the parameter count and effective rank selection.
  • regularization strengths for smoothness and sparsity
    Hyperparameters controlling the spatial regularization of the orthonormal basis; fitted or tuned during the ADMM procedure.
axioms (2)
  • domain assumption Residual field exhibits sufficient spatial smoothness and sparsity to make the orthonormal basis identifiable
    Invoked to justify turning a generic low-rank factorization into a structured spatial representation.
  • domain assumption Mini-batch ADMM converges to a useful solution for the joint basis and score estimation
    Required for the tractable learning procedure described.

pith-pipeline@v0.9.0 · 5551 in / 1639 out tokens · 40461 ms · 2026-05-13T01:35:14.129781+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

Reference graph

Works this paper leans on

65 extracted references · 65 canonical work pages · 2 internal anchors

  1. [1]

    Fixed Rank Kriging for Very Large Spatial Data Sets , journal =

    Noel Cressie and G. Fixed Rank Kriging for Very Large Spatial Data Sets , journal =. 2008 , doi =

    work page 2008

  2. [2]

    Journal of Computational and Graphical Statistics , volume =

    Wen-Ting Wang and Hsin-Cheng Huang , title =. Journal of Computational and Graphical Statistics , volume =. 2017 , doi =

    work page 2017

  3. [3]

    Technometrics , volume =

    ShengLi Tzeng and Hsin-Cheng Huang , title =. Technometrics , volume =. 2018 , doi =

    work page 2018

  4. [4]

    Reich and Ying Sun , title =

    Wanfang Chen and Yuxiao Li and Brian J. Reich and Ying Sun , title =. Statistica Sinica , volume =. 2024 , doi =

    work page 2024

  5. [5]

    Stat , volume =

    Ding-Chih Lin and Hsin-Cheng Huang and Sheng Li Tzeng , title =. Stat , volume =. 2023 , doi =

    work page 2023

  6. [6]

    Journal of the Royal Statistical Society: Series B (Statistical Methodology) , volume=

    Probabilistic principal component analysis , author=. Journal of the Royal Statistical Society: Series B (Statistical Methodology) , volume=. 1999 , publisher=

    work page 1999

  7. [7]

    Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining , pages=

    Optuna: A Next-generation Hyperparameter Optimization Framework , author=. Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining , pages=. 2019 , organization=

    work page 2019

  8. [8]

    Choi, Sang Won and Kim, Brian H. S. , title =. Sustainability , volume =. 2021 , doi =

    work page 2021

  9. [9]

    5 concentration based on deep learning and mode decomposition methods , author=

    A short-term prediction model of PM2. 5 concentration based on deep learning and mode decomposition methods , author=. Applied Sciences , volume=. 2021 , publisher=

    work page 2021

  10. [10]

    5 concentration in Ningxia Hui autonomous region based on PCA-attention-LSTM , author=

    Prediction of PM2. 5 concentration in Ningxia Hui autonomous region based on PCA-attention-LSTM , author=. Atmosphere , volume=. 2022 , publisher=

    work page 2022

  11. [11]

    and Jolliffe, I

    Hannachi, A. and Jolliffe, I. T. and Stephenson, D. B. , title =. International Journal of Climatology , volume =. 2007 , doi =

    work page 2007

  12. [12]

    Annual Review of Statistics and Its Application , volume=

    Statistical deep learning for spatial and spatiotemporal data , author=. Annual Review of Statistics and Its Application , volume=. 2023 , publisher=

    work page 2023

  13. [13]

    Journal of Computational and Graphical Statistics , pages=

    Efficient Large-scale Nonstationary Spatial Covariance Function Estimation Using Convolutional Neural Networks , author=. Journal of Computational and Graphical Statistics , pages=. 2024 , publisher=

    work page 2024

  14. [14]

    arXiv preprint arXiv:2410.04312 , year=

    Adjusting for Spatial Correlation in Machine and Deep Learning , author=. arXiv preprint arXiv:2410.04312 , year=

    work page arXiv

  15. [15]

    1993 , publisher=

    Nonparametric regression and generalized linear models: a roughness penalty approach , author=. 1993 , publisher=

    work page 1993

  16. [16]

    Gaussian Error Linear Units (GELUs)

    Gaussian error linear units (gelus) , author=. arXiv preprint arXiv:1606.08415 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  17. [17]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Deep residual learning for image recognition , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    work page

  18. [18]

    Advances in neural information processing systems , volume=

    Revisiting deep learning models for tabular data , author=. Advances in neural information processing systems , volume=

    work page

  19. [19]

    Journal of the Royal Statistical Society: Series B (Methodological) , volume=

    Regression shrinkage and selection via the lasso , author=. Journal of the Royal Statistical Society: Series B (Methodological) , volume=. 1996 , publisher=

    work page 1996

  20. [20]

    Exponential moving average of weights in deep learning: Dynamics and benefits,

    Exponential moving average of weights in deep learning: Dynamics and benefits , author=. arXiv preprint arXiv:2411.18704 , year=

    work page arXiv

  21. [21]

    Averaging weights leads to wider optima and better generalization.arXiv preprint arXiv:1803.05407, 2018

    Averaging weights leads to wider optima and better generalization , author=. arXiv preprint arXiv:1803.05407 , year=

    work page arXiv

  22. [22]

    5 datasets in the Beijing--Tianjin--Hebei region from 2015 to 2021 using VIIRS/DNB and deep learning model , author=

    High spatial resolution nighttime PM2. 5 datasets in the Beijing--Tianjin--Hebei region from 2015 to 2021 using VIIRS/DNB and deep learning model , author=. Remote Sensing , volume=. 2023 , publisher=

    work page 2015

  23. [23]

    5 concentration in the Beijing-Tianjin-Hebei region using a random forest model with a 0.01 0.01 spatial resolution , author=

    Estimating the daily PM2. 5 concentration in the Beijing-Tianjin-Hebei region using a random forest model with a 0.01 0.01 spatial resolution , author=. Environment international , volume=. 2020 , publisher=

    work page 2020

  24. [24]

    Plant Phenomics , volume=

    Global Wheat Head Detection (GWHD) dataset: a large and diverse dataset of high-resolution RGB-labelled images to develop and benchmark wheat head detection methods , author=. Plant Phenomics , volume=. 2020 , publisher=

    work page 2020

  25. [25]

    2021 , eprint=

    Global Wheat Head Dataset 2021: an update to improve the benchmarking wheat head localization with more diversity , author=. 2021 , eprint=

    work page 2021

  26. [26]

    Proceedings of the 2011 31st International Conference on Distributed Computing Systems Workshops (ICDCSW) , year =

    Finding a “Kneedle” in a Haystack: Detecting Knee Points in System Behavior , author =. Proceedings of the 2011 31st International Conference on Distributed Computing Systems Workshops (ICDCSW) , year =

    work page 2011

  27. [27]

    , title =

    Cattell, Raymond B. , title =. Psychometrika , year =

    work page

  28. [28]

    Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences , volume=

    Cautionary tales on air-quality improvement in Beijing , author=. Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences , volume=. 2017 , publisher=

    work page 2017

  29. [29]

    Bishop , title =

    Christopher M. Bishop , title =. 2006 , publisher =

    work page 2006

  30. [30]

    2009 , publisher =

    Trevor Hastie and Robert Tibshirani and Jerome Friedman , title =. 2009 , publisher =

    work page 2009

  31. [31]

    5 concentrations in Beijing using CNN-Bi LSTM with CBAM , author=

    Prediction of multi-site PM2. 5 concentrations in Beijing using CNN-Bi LSTM with CBAM , author=. Atmosphere , volume=. 2022 , publisher=

    work page 2022

  32. [32]

    Expert Systems with Applications , volume=

    Multi-hour and multi-site air quality index forecasting in Beijing using CNN, LSTM, CNN-LSTM, and spatiotemporal clustering , author=. Expert Systems with Applications , volume=. 2021 , publisher=

    work page 2021

  33. [33]

    Artificial intelligence and machine learning for multi-domain operations applications , volume=

    Super-convergence: Very fast training of neural networks using large learning rates , author=. Artificial intelligence and machine learning for multi-domain operations applications , volume=. 2019 , organization=

    work page 2019

  34. [34]

    7th Int , author=

    Decoupled weight decay regularization. 7th Int , author=. Conf. Learn. Represent. ICLR , year=

    work page

  35. [35]

    Adam: A Method for Stochastic Optimization

    Adam: A method for stochastic optimization , author=. arXiv preprint arXiv:1412.6980 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  36. [36]

    2017 , publisher=

    Generalized Additive Models: An Introduction with R , author=. 2017 , publisher=

    work page 2017

  37. [37]

    Journal of the American Statistical Association , volume=

    Spatial modeling with spatially varying coefficient processes , author=. Journal of the American Statistical Association , volume=. 2003 , publisher=

    work page 2003

  38. [38]

    Foundations and Trends in Optimization , year =

    Parikh, Neal and Boyd, Stephen , title =. Foundations and Trends in Optimization , year =

    work page

  39. [39]

    and Mahony, Robert and Sepulchre, Rodolphe , title =

    Absil, P.-A. and Mahony, Robert and Sepulchre, Rodolphe , title =. 2008 , isbn =

    work page 2008

  40. [40]

    Journal of Machine Learning Research , year =

    Journée, Mathieu and Nesterov, Yurii and Richtárik, Peter and Sepulchre, Rodolphe , title =. Journal of Machine Learning Research , year =

    work page

  41. [41]

    Journal of Computational and Graphical Statistics , year =

    Zou, Hui and Hastie, Trevor and Tibshirani, Robert , title =. Journal of Computational and Graphical Statistics , year =

    work page

  42. [42]

    and Van Loan, Charles F

    Golub, Gene H. and Van Loan, Charles F. , title =. 2013 , isbn =

    work page 2013

  43. [43]

    Proceedings of the 30th International Conference on Machine Learning (ICML) , year =

    Van Vu and Jian Lei and Karl Rohe , title =. Proceedings of the 30th International Conference on Machine Learning (ICML) , year =

    work page

  44. [44]

    Journal of Machine Learning Research , year =

    Journee, Mathieu and Nesterov, Yurii and Richtarik, Peter and Sepulchre, Rodolphe , title =. Journal of Machine Learning Research , year =

    work page

  45. [45]

    Foundations and Trends

    Distributed optimization and statistical learning via the alternating direction method of multipliers , author=. Foundations and Trends. 2011 , publisher=

    work page 2011

  46. [46]

    Geoscientific Model Development , volume =

    M. Geoscientific Model Development , volume =. 2022 , doi =

    work page 2022

  47. [47]

    and Lo, Wan-Yen and Doll

    Kirillov, Alexander and Mintun, Eric and Ravi, Nikhila and Mao, Hanzi and Rolland, Chloe and Gustafson, Laura and Xiao, Tete and Whitehead, Spencer and Berg, Alexander C. and Lo, Wan-Yen and Doll. Segment Anything , booktitle =

    work page

  48. [48]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

    Liu, Zhuang and Mao, Hanzi and Wu, Chao-Yuan and Feichtenhofer, Christoph and Darrell, Trevor and Xie, Saining , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

    work page

  49. [49]

    International Conference on Learning Representations (ICLR) , year =

    Dosovitskiy, Alexey and Beyer, Lucas and Kolesnikov, Alexander and Weissenborn, Dirk and Zhai, Xiaohua and Unterthiner, Thomas and Dehghani, Mostafa and Minderer, Matthias and Heigold, Georg and Gelly, Sylvain and Uszkoreit, Jakob and Houlsby, Neil , title =. International Conference on Learning Representations (ICLR) , year =

    work page

  50. [50]

    Hu and Yelong Shen and Phillip Wallis and Zeyuan Allen-Zhu and Yuanzhi Li and Shean Wang and Lu Wang and Weizhu Chen , title =

    Edward J. Hu and Yelong Shen and Phillip Wallis and Zeyuan Allen-Zhu and Yuanzhi Li and Shean Wang and Lu Wang and Weizhu Chen , title =. International Conference on Learning Representations (ICLR) , year =

    work page

  51. [51]

    Parameter-Efficient Transfer Learning for

    Neil Houlsby and Andrei Giurgiu and Stanis. Parameter-Efficient Transfer Learning for. International Conference on Machine Learning (ICML) , year =

    work page

  52. [52]

    Emmanuel J. Cand. Enhancing Sparsity by Reweighted _1 Minimization , journal =

    work page

  53. [53]

    Coull , title =

    Alan Agresti and Brent A. Coull , title =. The American Statistician , volume =

    work page

  54. [54]

    Multinomial Logistic Regression Algorithm , journal =

    Dankmar B. Multinomial Logistic Regression Algorithm , journal =. 1992 , doi =

    work page 1992

  55. [55]

    Jaakkola and Michael I

    Tommi S. Jaakkola and Michael I. Jordan , title =. Statistics and Computing , volume =. 2000 , doi =

    work page 2000

  56. [56]

    2005 , doi =

    Vladimir Vovk and Alexander Gammerman and Glenn Shafer , title =. 2005 , doi =

    work page 2005

  57. [57]

    Transactions on Machine Learning Research , year =

    Parameter-Efficient Fine-Tuning for Large Models: A Comprehensive Survey , author =. Transactions on Machine Learning Research , year =

    work page

  58. [58]

    2024 , publisher =

    Liu, Shih-Yang and Wang, Chien-Yi and Yin, Hongxu and Molchanov, Pavlo and Wang, Yu-Chiang Frank and Cheng, Kwang-Ting and Chen, Min-Hung , booktitle =. 2024 , publisher =

    work page 2024

  59. [59]

    Localized

    Barazandeh, Babak and Majumdar, Subhabrata and Rajyaguru, Om and Michailidis, George , journal =. Localized. 2025 , doi =

    work page 2025

  60. [60]

    International Conference on Learning Representations , year =

    Bayesian Low-rank Adaptation for Large Language Models , author =. International Conference on Learning Representations , year =

    work page

  61. [61]

    and Wang, Hao , booktitle =

    Wang, Yibin and Shi, Haizhou and Han, Ligong and Metaxas, Dimitris N. and Wang, Hao , booktitle =. 2024 , url =

    work page 2024

  62. [62]

    Minimal Ranks, Maximum Confidence: Parameter-efficient Uncertainty Quantification for

    Marsza. Minimal Ranks, Maximum Confidence: Parameter-efficient Uncertainty Quantification for. Findings of the Association for Computational Linguistics: EMNLP 2025 , pages =. 2025 , publisher =

    work page 2025

  63. [63]

    and Soden, Brian J

    Klein, Stephen A. and Soden, Brian J. and Lau, Ngar-Cheung , journal =. Remote Sea Surface Temperature Variations during

    work page

  64. [64]

    Impact of the

    Yang, Jianling and Liu, Qinyu and Xie, Shang-Ping and Liu, Zhengyu and Wu, Lixin , journal =. Impact of the

    work page

  65. [65]

    Xie, Shang-Ping and Hu, Kaiming and Hafner, Jan and Tokinaga, Hiroki and Du, Yan and Huang, Gang and Sampe, Takeaki , journal =

    work page