Pith. sign in

REVIEW 5 major objections 5 minor 43 references

The paper establishes that, under sudden concept drift, the optimal retraining policy is a single-switch bang-bang control — front-load training when concept durations have decreasing mean residual life, defer it when they have increasing m

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

2026-08-03 16:33 UTC pith:EUONGVP4

load-bearing objection Novel and worth refereeing, but the central DMRL/IMRL and deployment proofs as written don't hold up. the 5 major comments →

arxiv 2512.12816 v2 pith:EUONGVP4 submitted 2025-12-14 cs.LG cs.NI

Optimal Resource Allocation for ML Model Training and Deployment under Concept Drift

Hasan Burhan Beytur , Haris Vikalo , Kevin S Chan , Gustavo de Veciana This is my paper
classification cs.LG cs.NI
keywords concept driftresource allocationmean residual lifeDMRLIMRLbang-bang controldeployment schedulingrenewal-reward theorem
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 tries to establish that the right way to spend a fixed training budget under sudden concept drift is not a heuristic but a provably optimal switching policy, and that the direction of the switch is decided by one statistical property of the concept duration distribution. When concepts have decreasing mean residual life (DMRL) — the expected remaining time until the next drift shrinks as a concept ages — the optimal policy is front-loading: train at full capacity at the start of each concept, then stop. When concepts have increasing mean residual life (IMRL), the optimum delays training instead. The paper also claims that the separate problem of when to push updated models to clients is quasi-convex under mild conditions, and that a randomized scheduler can exactly meet a deployment-rate limit while nearly matching optimal client-side loss. If this is right, a provider only needs to classify the aging behavior of its concept durations to choose a near-optimal retraining and deployment strategy.

Core claim

The central claim is Theorem 3.4: for any convex, decreasing expected loss function g-bar and any positive budget, the optimal resource allocation is a single-switch front-loading control — train at maximum rate until a switching time, then idle — if and only if the concept duration Y has decreasing mean residual life. The switching time is pinned by the budget through the survival function of Y. Theorem 3.5 and Corollary 3.6 say that under IMRL the optimum instead idles at the start of each concept, and with a linearly decreasing loss curve the optimal policy is single-switch back-loading. On the deployment side, the paper proves the client-side loss minimization is quasi-convex under mild

What carries the argument

The central object is the mean residual life function m_Y(t) = E[Y - t | Y > t]; DMRL means this function is non-increasing, IMRL means it is non-decreasing. It enters through the switching function of Pontryagin's maximum principle applied to a renewal-reward reduction of the time-average loss: the survival function of the concept duration weights the objective, and the monotonicity of m_Y determines whether the switching function crosses zero once, which selects front-loading versus back-loading. On the deployment side, the objective is written as an integral of the expected loss at the most recent deployment offset weighted by the survival function, and that representation is what makes t

Load-bearing premise

The load-bearing premise is that after every concept change both the server-side and client-side expected losses reset to a common initial level, independent of the previous model; if the old model leaves residual knowledge, the per-concept cycles are no longer identical and the DMRL/IMRL optimality proofs no longer directly apply.

What would settle it

Construct a concept duration distribution whose mean residual life is neither monotone (for example, a bathtub-hazard mixture of early failures and late wear-out), choose a convex decreasing expected loss function and a binding budget, and numerically solve the Pontryagin optimality conditions; if the optimal allocation is not single-switch front-loading for some such instance, the 'if and only if' claim of Theorem 3.4 is false. Alternatively, measure the starting loss after each drift event in a real deployment log: significant dependence on how long the previous concept was trained would con

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • For DMRL concept durations, the optimal use of a limited budget is to spend it all at the start of each concept; the paper's simulations show up to 71.8 percent reduction in time-average loss compared with a fixed-rate allocation under the same budget.
  • For IMRL durations, the optimal policy initially idles and then trains; fixed or periodic retraining schedules are provably suboptimal in this regime.
  • A provider does not need to know the full expected-loss curve to choose the structure of the optimal policy — only the aging class (DMRL or IMRL) of the concept duration distribution, plus the survival function and budget to set the switch time.
  • The optimal deployment schedule concentrates updates near the start of a concept, and for exponential durations the inter-deployment intervals have a closed-form backward chain.
  • The randomized deployment scheduler achieves full utilization of the deployment budget without relying on convexity of the survival function, making it a computationally simple near-optimal alternative for non-convex cases.
  • The paper's Erlang-2 simulation shows a concrete case where the randomized scheduler is strictly worse than the optimal deterministic schedule, so the near-optimality guarantee is not universal.

Where Pith is reading between the lines

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

  • Beyond the paper's reset assumption: if a model trained on the previous concept transfers even partial knowledge to the new concept, the per-concept starting loss depends on how much training came before, so the cycles are no longer i.i.d. and the DMRL/IMRL policy shapes would need to be re-derived with state-dependent reset levels.
  • A natural testable extension is to measure post-drift initial loss as a function of training investment in the prior concept; if the starting loss decreases with prior training time, the renewal-reward simplification is violated and the optimal policy may shift toward continuous retraining.
  • The same survival-function-weighted machinery could be applied to systems where drift is detected rather than observed, with the 'concept duration' replaced by the detected inter-change time; the aging classification would then depend on the detector's operating characteristics.
  • The chain-structured deployment schedule implies a direct empirical check: in a real deployment log, intervals between updates should shrink as time since the last drift grows; if they do not, the assumptions on the loss curve or duration distribution may not hold.

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

5 major / 5 minor

Summary. The paper introduces a renewal-theoretic framework for allocating training resources and scheduling model deployments under sudden concept drift. It models the server-side expected loss as a function of cumulative training effort, derives time-average objectives under static policies, and uses Pontryagin's Maximum Principle to claim that optimal training policies are bang-bang: front-loading if concept durations have DMRL (Theorem 3.4) and back-loading/idling if they have IMRL (Theorem 3.5, Corollary 3.6). For deployment, it claims the optimization problem is quasi-convex (Lemma 4.2), characterizes optimal deterministic schedulers (Theorem 4.3), and proposes a randomized scheduler that exactly meets the deployment rate constraint (Theorem 4.5). Several appendices provide proofs.

Significance. If correct, the paper would give a model-agnostic, practically actionable characterization: a provider needs only know the aging class of concept durations to choose an optimal retraining/deployment policy. The modeling contribution is clean and the renewal-reward reduction to infinite-horizon optimal control is natural. The paper also includes simulation evidence and does not hide fitted parameters in the theorems. However, the central proof apparatus has multiple load-bearing errors: the IMRL back-loading proof reverses a sign and uses the opposite inequality from Definition 3.3; the 'if and only if' in Theorem 3.4 has only one direction proved; the quasi-convexity proof of Lemma 4.2 is invalid; and the monotonicity and rate-saturation arguments behind Theorems 4.3 and 4.5 are not established. As a result, the main theoretical claims are currently unsupported.

major comments (5)
  1. [Appendix A.3, Eqs. (54)-(58); Definition 3.3] The proof of Theorem 3.5 is internally inconsistent. If e*(t)=0 on [0,t*) and φ(t*)=0, optimality requires φ(t)>0 for t<t*, so near t* the function must decrease to zero; hence φ'(t) should be ≤0, not 'should be positive' as claimed. Moreover, the chain 'if m_Y(t)>m_Y(t*) for t<t* ... 1-m_Y(t)h_Y(t)>0, where the last inequality follows from the IMRL property' is reversed: IMRL means m_Y is nondecreasing, so m_Y(t)≤m_Y(t*) for t<t*, and IMRL gives h_Y(t)m_Y(t)≥1, i.e. 1-h_Y(t)m_Y(t)≤0. The inequality used is a DMRL ordering. Corollary 3.6 (Appendix A.4, Eqs. (64)-(65)) repeats the same error. Thus Theorems 3.5 and 3.6 are not proven, and the claimed sign pattern would actually prescribe e=M on [0,t*) by (16).
  2. [Theorem 3.4 and Appendix A.2] The theorem is stated as 'if and only if', but the proof only gives the sufficiency direction under DMRL. It constructs the switch time from the budget and verifies the sign pattern of φ(t) assuming m_Y is decreasing. There is no argument for the converse: that optimality of front-loading for every convex decreasing g-bar and every budget B forces Y to be DMRL. The missing direction is load-bearing for the paper's classification claim that the aging class fully determines the policy structure.
  3. [Lemma 4.2 and Appendix A.5, Eqs. (70)-(74)] The quasi-convexity proof is invalid. The pointwise inequality g-bar(s^{(z)}(t)) F̄_Y(t) ≤ max(g-bar(s^{(1)}(t))F̄_Y(t), g-bar(s^{(2)}(t))F̄_Y(t)) does not imply the integral inequality h(z) ≤ max(h(δ^{(1)}), h(δ^{(2)})), because the integral of a pointwise maximum is not the maximum of the integrals. The proof then invokes an extra 'component-wise greater or equal' assumption on the two offset sets, which is not part of the lemma statement and does not cover arbitrary pairs. Consequently the quasi-convexity of Problem (32) and the KKT-sufficiency results that rely on it (Theorems 4.3 and 4.5) are unsupported.
  4. [Theorem 4.3 and Appendix A.6, Eqs. (80)-(84)] The claim that the effective deployment rate r_e is monotonically increasing in N_D is essential for Theorem 4.5, but it is not proven. The sentence 'By the convexity of g-bar, the monotonicity of F̄_Y, and the KKT conditions, one can show that δ_j^*(N_D+1)<δ_j^*(N_D)' is a nontrivial comparative-statics assertion with no derivation. Eq. (82)-(84) then subtract terms for the same index j across two different schedules without accounting for the fact that the optimal offsets are not the same, so the claimed positivity does not follow. Theorem 4.5's existence of N_D^* and γ therefore lacks a valid basis.
  5. [Theorem 4.5 and Appendix A.8; Abstract] The theorem assumes the optimal scheduler for Problem (32) has r_e({δ_k^*}) = r_D E[Y], i.e., that the rate constraint is exactly saturated. Problem (32)'s constraint is an inequality, and no argument rules out a slack optimal solution. The proof also does not establish that a convex combination of rates for consecutive N_D can match an arbitrary saturation level without additional density/coverage arguments. Finally, the abstract and Section 4 claim 'near-optimal client-side performance' for the randomized scheduler, but Theorem 4.5 only proves exact rate matching; no bound on the loss suboptimality is given.
minor comments (5)
  1. [Appendix A.6, Eq. (38)] Typo: 'm_Y(d*_{N_D})' should read 'm_Y(δ*_{N_D})'.
  2. [Lemma 3.2 and Appendix A.1] The lemma states existence of an optimal solution, but the proof only demonstrates convexity of J and affine-ness of C. Existence also requires compactness or coercivity of the feasible set; this should be stated and proved.
  3. [Section 2, bullet list] The assumption that 'at each concept change, the expected concept loss of both the deployed and server-side models resets to a common initial level' is very restrictive. In practice a model trained on the previous concept may transfer knowledge to the new concept. The authors should at least discuss the limitation and the extent to which the renewal-cycle analysis depends on it.
  4. [Appendix A.6, Eq. (78)] The derivative ∂h/∂δ_j appears to have the integration limits swapped; the integral multiplying g-bar'(δ_j^*) should be ∫_{δ_j^*}^{δ_{j+1}^*} F̄_Y(t)dt, not ∫_{δ_{j-1}^*}^{δ_j^*} F̄_Y(t)dt, to match the expression in (38).
  5. [Abstract and Section 4] The phrase 'provably suboptimal' under IMRL is not supported by any theorem: Theorem 3.5 only claims idling before training, not suboptimality of fixed schedules. Likewise, 'near-optimal client-side performance' is not a proven theorem statement but a simulation observation.

Circularity Check

0 steps flagged

No significant circularity; the derivations are self-contained from stated assumptions.

full rationale

The paper's training-resource theorems (Theorem 3.4, Theorem 3.5, Corollary 3.6) are derived from explicitly stated modeling assumptions via the renewal-reward theorem and Pontryagin's Maximum Principle, with external references for those standard tools. The DMRL/IMRL classification is defined independently of the optimal-policy conclusions; the proofs attempt to show how the sign of the switching function follows from the MRL monotonicity, so the policy structure is not assumed into the theorems. The deployment results are also derived from the stated loss and survival-function assumptions rather than fitted to data. There are no self-citations, no parameter fitted to a subset of data and then renamed a prediction, and no ansatz imported from the authors' prior work. Theorem 4.5's rate-matching equation defines the randomization probability so that the effective rate equals the limit; this is a construction rather than a hidden circular derivation, and the claimed near-optimal client-side performance is supported by simulations, not by that tautological rate equation. The reviewer's concerns about the internal validity of the IMRL proof (e.g., sign of the switching-function derivative) are potential correctness errors, not circularity. Accordingly, no specific circular step is exhibited, and the score is 0.

Axiom & Free-Parameter Ledger

0 free parameters · 7 axioms · 0 invented entities

The paper assumes a stylized world: i.i.d concept durations, convex decreasing expected-loss curves, linear resource scaling, reset losses, static policies, and known distributions. These are not fitted, so no free parameters are listed; but they are domain assumptions that limit practical reach.

axioms (7)
  • domain assumption Concept durations Y_i are i.i.d., have finite mean, and are independent of the loss curves G_i.
    Used to invoke the Renewal-reward Theorem (Eq. 5) and to make the infinite-horizon time-average objective well-defined; if durations are autocorrelated or dependent on loss curves, the objective separates differently.
  • domain assumption Each concept's expected loss curve g_i is convex and decreasing; the aggregate ḡ inherits convexity and decreasingness.
    Convexity is needed for Lemma 3.2 (convex functional) and for the bang-bang/single-crossing arguments; the paper notes individual g_i convexity can be relaxed but ḡ must be convex.
  • domain assumption Training speed scales linearly with resource allocation e(t); loss depends only on the cumulative resource ∫e dτ.
    This horizontal-scaling transformation (Eq. 1) is load-bearing: without linearity in e, the control problem is not linear in the control and the bang-bang structure fails.
  • domain assumption At each concept change, expected concept loss of both deployed and server-side models resets to a common initial level.
    Bullet in Section 2; used to make each renewal cycle statistically identical. If pre-drift model quality carries over, optimal policies may differ.
  • domain assumption Only static (concept-independent) resource-allocation and deployment policies are considered.
    Sections 3 and 4 restrict attention to static policies; the 'optimal' claims are within this restricted class, not over all history-dependent policies.
  • domain assumption The distribution of Y and the expected loss curve ḡ are known to the provider.
    Section 4 requires F_Y and ḡ; no estimation method or robustness analysis is given.
  • standard math Pontryagin Maximum Principle, Arrow sufficiency, Renewal-reward theorem, KKT conditions, and Mean Value Theorem apply as used.
    Used in Appendix A as proof tools; assumes infinite-horizon control regularity (free terminal state, bounded controls).

pith-pipeline@v1.3.0-alltime-deepseek · 22848 in / 17522 out tokens · 153259 ms · 2026-08-03T16:33:01.263067+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Optimal Resource Allocation for ML Model Training and Deployment under Concept Drift." pith.science (2026). https://pith.science/paper/EUONGVP4

@misc{pith2026251212816,
  author       = {Pith},
  title        = {Pith review of: Optimal Resource Allocation for ML Model Training and Deployment under Concept Drift},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EUONGVP4}},
  note         = {Machine review of arXiv:2512.12816}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

We study how to allocate resources for training and deployment of machine learning (ML) models under concept drift and limited budgets. We consider a setting in which a model provider distributes trained models to multiple clients whose devices support local inference but lack the ability to retrain those models, placing the burden of performance maintenance on the provider. We introduce a model-agnostic framework that captures the interaction between resource allocation, concept drift dynamics, and deployment timing. We show that optimal training policies depend critically on the aging properties of concept durations. Under sudden concept changes, we derive optimal training policies subject to budget constraints when concept durations follow distributions with Decreasing Mean Residual Life (DMRL), and show that intuitive heuristics are provably suboptimal under Increasing Mean Residual Life (IMRL). We further study model deployment under communication constraints, prove that the associated optimization problem is quasi-convex under mild conditions, and propose a randomized scheduling strategy that achieves near-optimal client-side performance. These results offer theoretical and algorithmic foundations for cost-efficient ML model management under concept drift, with implications for continual learning, distributed inference, and adaptive ML systems.

Figures

Figures reproduced from arXiv: 2512.12816 by Gustavo de Veciana, Haris Vikalo, Hasan Burhan Beytur, Kevin S Chan.

Figure 1
Figure 1. Figure 1: Top: training resource allocation 𝑒𝑖(𝑡). Bottom: sample paths of the expected concept loss for the server L (𝑡) and client models L (ˆ 𝑡), as defined in (1) and (2). Shaded regions indicate fluctuations in observed evaluation loss. 𝑇𝑖 denotes concept change times, and 𝐷𝑖,𝑗 denote deployment times. with 𝑒𝑖(𝑡), which we model through a horizontal scaling transformation of 𝐺𝑖(·) controlled by the allocation f… view at source ↗
Figure 2
Figure 2. Figure 2: Comparison of optimal, fixed and delayed resource allocation policies under sudden concept drift. All [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of periodic, optimal, and randomized deployment policies under different concept duration [PITH_FULL_IMAGE:figures/full_fig_p015_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of optimal and fixed resource allocation policies under sudden concept drift for different [PITH_FULL_IMAGE:figures/full_fig_p024_4.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

43 extracted references · 5 canonical work pages

  1. [1]

    Lucas Baier, Vincent Kellner, Niklas Kühl, and Gerhard Satzger. 2020. Switching Scheme: A Novel Approach for Handling Incremental Concept Drift in Real-World Data Sets. arXiv:2011.02738 [cs.LG] https://arxiv.org/abs/2011.02738

    Pith/arXiv arXiv 2020

  2. [2]

    Jihwan Bang, Hyunseo Koh, Seulki Park, Hwanjun Song, Jung-Woo Ha, and Jonghyun Choi. 2022. Online continual learning on a contaminated data stream with blurry task boundaries.Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition(2022), 9275–9284

    2022

  3. [3]

    Barlow and Frank Proschan

    Richard E. Barlow and Frank Proschan. 1974/1975. Statistical Theory of Reliability and Life Testing: Probability Models. InStatistical Theory of Reliability and Life Testing: Probability Models. Holt, Rinehart and Winston, New York

    1974

  4. [4]

    Bo Bergman and Bengt Klefsjö. 1989. A Family of Test Statistics for Detecting Monotone Mean Residual Life.Journal of Statistical Planning and Inference21, 2 (Feb. 1989), 161–178. doi:10.1016/0378-3758(89)90002-5

    work page doi:10.1016/0378-3758(89)90002-5 1989

  5. [5]

    Albert Bifet and Ricard Gavaldà. 2007. Learning from Time-Changing Data with Adaptive Windowing. InProceedings of the 2007 SIAM International Conference on Data Mining (SDM). Society for Industrial and Applied Mathematics, 443–448. doi:10.1137/1.9781611972771.42

    doi:10.1137/1.9781611972771.42 2007

  6. [6]

    Budka and B

    M. Budka and B. Gabrys. 2018. Change-point Detection in Evolving Data Streams Using Ensembles of Classifiers.IEEE Signal Processing Letters25, 9 (2018), 1353–1357. doi:10.1109/LSP.2018.2849385

    arXiv 2018

  7. [7]

    D. Chen, S. Yang, and J. Li. 2024. What Role Do Small Models Play in a World of Giants?. InInternational Conference on Learning Representations

    2024

  8. [8]

    Weihao Cheng, Sarah Erfani, Rui Zhang, and Ramamohanarao Kotagiri. 2018. Learning datum-wise sampling frequency for energy-efficient human activity recognition. 32, 1 (2018)

    2018

  9. [9]

    Ben Cottier, Robi Rahman, Loredana Fattorini, Nestor Maslej, Tamay Besiroglu, and David Owen. 2025. The rising costs of training frontier AI models. arXiv:2405.21015 [cs.CY] https://arxiv.org/abs/2405.21015

    Pith/arXiv arXiv 2025

  10. [10]

    Domhan, J

    T. Domhan, J. Springenberg, and F. Hutter. 2015. Speeding up Automatic Hyperparameter Optimization of Deep Neural Networks by Extrapolation of Learning Curves. InProceedings of the 24th International Conference on Artificial Intelligence. 3460–3468

    2015

  11. [11]

    Gallager

    Robert G. Gallager. 1996.Discrete Stochastic Processes. Springer US, Boston, MA. doi:10.1007/978-1-4615-2329-1

    work page doi:10.1007/978-1-4615-2329-1 1996

  12. [12]

    J. Gama, I. Žliobait˙e, A. Bifet, M. Pechenizkiy, and A. Bouchachia. 2014. A Survey on Concept Drift Adaptation.Comput. Surveys46, 4 (2014), 44:1–44:37. doi:10.1145/2523813

    doi:10.1145/2523813 2014

  13. [13]

    W. J. Hall and Jon A. Wellner. 2020. Estimation of Mean Residual Life. InStatistical Modeling for Biological Systems: In Memory of Andrei Yakovlev, Anthony Almudevar, David Oakes, and Jack Hall (Eds.). Springer International Publishing, Cham, 169–189. doi:10.1007/978-3-030-34675-1_10

    work page doi:10.1007/978-3-030-34675-1_10 2020

  14. [14]

    Meng Han, Zhiqiang Chen, Muhang Li, Hongxin Wu, and Xilong Zhang. 2022. A survey of active and passive concept drift handling methods.IEEE Transactions on Knowledge and Data Engineering38, 4 (2022), 1492–1535. doi:10.1111/coin.12520

    doi:10.1111/coin.12520 2022

  15. [15]

    Hoffmann, S

    J. Hoffmann, S. Borgeaud, A. Mensch, P. Buchlovsky, T. Cai, E. Rutherford, K. Millican, C. Jones, B. Bos, S. Gray, C. Leahy, E. Conway, Z. Dai, A. Mirhoseini, and E. Grefenstette. 2022. Training Compute-Optimal Large Language Models. arXiv preprint arXiv:2203.15556(2022)

    Pith/arXiv arXiv 2022

  16. [16]

    Weiqiang Huang, Juecen Zhan, Yumeng Sun, Xu Han, Tai An, and Nan Jiang. 2025. Context-Aware Adaptive Sampling for Intelligent Data Acquisition Systems Using DQN.arXiv preprint arXiv:2504.09344(2025)

    Pith/arXiv arXiv 2025

  17. [17]

    Hussein and N

    R. Hussein and N. Gupta. 2025. ChatGPT’s Impact Across Sectors: A Survey.International Journal of Human–Computer Studies152 (2025), 102761. doi:10.1016/j.ijhcs.2024.102761

    arXiv 2025

  18. [18]

    Kaplan, S

    J. Kaplan, S. McCandlish, T. Henighan, T. Brown, B. Chess, R. Child, S. Gray, A. Radford, J. Wu, and D. Amodei. 2020. Scaling Laws for Neural Language Models.arXiv preprint arXiv:2001.08361(2020)

    Pith/arXiv arXiv 2020

  19. [19]

    Kim and S

    H. Kim and S. Park. 2025. Tools for Understanding How Large Language Models Work.arXiv preprint arXiv:2501.01234 (2025)

    Pith/arXiv arXiv 2025

  20. [20]

    Kochar, Hari Mukerjee, and Francisco J

    Subhash C. Kochar, Hari Mukerjee, and Francisco J. Samaniego. 2000. Estimation of a Monotone Mean Residual Life. The Annals of Statistics28, 3 (2000), 905–921. jstor:2674059

    2000

  21. [21]

    J. Z. Kolter and M. A. Maloof. 2007. Dynamic Weighted Majority: An Ensemble Method for Drifting Concepts. In Proceedings of the 2007 IEEE International Conference on Data Mining. 123–132

    2007

  22. [22]

    Dominik Kreuzberger, Niklas Kühl, and Sebastian Hirschl. 2022. Machine Learning Operations (MLOps): Overview, Definition, and Architecture. arXiv:2205.02302 [cs.LG] https://arxiv.org/abs/2205.02302

    Pith/arXiv arXiv 2022

  23. [23]

    Mengtian Li, Ersin Yumer, and Deva Ramanan. 2019. Budgeted training: Rethinking deep neural network training under resource constraints.arXiv preprint arXiv:1905.04753(2019)

    Pith/arXiv arXiv 2019

  24. [24]

    Ahmed Ali Linkon, Mujiba Shaima, Md Shohail Uddin Sarker, Badruddowza, Norun Nabi, Md Nasir Uddin Rana, Sandip Kumar Ghosh, Mohammad Anisur Rahman, Hammed Esa, and Faiaz Rahat Chowdhury. 2024. Advancements and Applications of Generative Artificial Intelligence and Large Language Models on Business Management: A Comprehensive Review.Journal of Computer Sci...

    2024

  25. [25]

    Jie Lu, Anjin Liu, Fan Dong, Feng Gu, João Gama, and Guangquan Zhang. 2019. Learning under Concept Drift: A Review.IEEE Transactions on Knowledge and Data Engineering31, 12 (2019), 2346–2363. doi:10.1109/TKDE.2018.2876857 Conference Name: IEEE Transactions on Knowledge and Data Engineering

    arXiv 2019

  26. [26]

    Sandeep Madireddy, Prasanna Balaprakash, Philip Carns, Robert Latham, Glenn K Lockwood, Robert Ross, Shane Snyder, and Stefan M Wild. 2019. Adaptive learning for concept drift in application performance modeling. In Proceedings of the 48th International Conference on Parallel Processing. 1–11

    2019

  27. [27]

    Ananth Mahadevan and Michael Mathioudakis. 2023. Cost-effective retraining of machine learning models.arXiv preprint arXiv:2310.04216(2023)

    Pith/arXiv arXiv 2023

  28. [28]

    Ajay Mandlekar, Yuke Zhu, Animesh Garg, Jonathan Booher, Max Spero, Albert Tung, Julian Gao, John Emmons, Anchit Gupta, Emre Orbay, et al. 2018. Roboturk: A crowdsourcing platform for robotic skill learning through imitation. InConference on Robot Learning. PMLR, 879–893

    2018

  29. [29]

    Aryan Mokhtari, Shahin Shahrampour, Ali Jadbabaie, and Alejandro Ribeiro. 2016. Online optimization in dynamic environments: Improved regret rates for strongly convex problems. (2016), 7195–7201. doi:10.1109/CDC.2016.7799379

    arXiv 2016

  30. [30]

    Moreno-García, J

    F. Moreno-García, J. del Campo-Ávila, N. García-Pedrajas, and S. Ventura. 2020. Dynamic Ensemble Selection Based on Local Accuracy for Non-Stationary Environments.Data Mining and Knowledge Discovery34, 3 (2020), 736–770. doi:10.1007/s10618-019-00658-8

    work page doi:10.1007/s10618-019-00658-8 2020

  31. [31]

    Maksim Muravev, Brazhenko , Dmitry, Somenkova , Anzhela, Golovkov , Alexander, and Ilia and Sergunin. [n. d.]. MLOps Architecture as a Future of Machine Learning.Journal of Computer Information Systems0, 0 ([n. d.]), 1–13. doi:10.1080/08874417.2025.2483826

    arXiv 2025

  32. [32]

    J. Park, K. Lee, and M. Cho. 2024. LLamaDuo: Dual-Mode Inference for On-Device LLMs. InProceedings of EMNLP. ACL, 1123–1135

    2024

  33. [33]

    Aurick Qiao, Sang Keun Choe, Suhas Jayaram Subramanya, Willie Neiswanger, Qirong Ho, Hao Zhang, Gregory R Ganger, and Eric P Xing. 2021. Pollux: Co-adaptive cluster scheduling for goodput-optimized deep learning. In15th {USENIX}Symposium on Operating Systems Design and Implementation ({OSDI}21)

    2021

  34. [34]

    Florence Regol, Leo Schwinn, Kyle Sprague, Mark Coates, and Thomas Markovich. 2025. When to retrain a machine learning model. https://openreview.net/forum?id=iGX0lwpUYj

    2025

  35. [35]

    Suresh P. Sethi. 2021.Optimal Control Theory: Applications to Management Science and Economics. Springer International Publishing, Cham. doi:10.1007/978-3-030-91745-6

    work page doi:10.1007/978-3-030-91745-6 2021

  36. [36]

    Ya Shen, Gang Chen, Hui Ma, and Mengjie Zhang. 2024. Cost-Aware Dynamic Cloud Workflow Scheduling Using Self-attention and Evolutionary Reinforcement Learning. InInternational Conference on Service-Oriented Computing. Springer, 3–18

    2024

  37. [37]

    Andrea Simonetto, Aryan Mokhtari, Alec Koppel, Geert Leus, and Alejandro Ribeiro. 2016. A Class of Prediction- Correction Methods for Time-Varying Convex Optimization.IEEE Transactions on Signal Processing64, 17 (2016), 4576–4591. doi:10.1109/TSP.2016.2568161 77 citations (Crossref) [2023-12-17]

    arXiv 2016

  38. [38]

    Jasper Stone, Raj Patel, Farbod Ghiasi, Sudip Mittal, and Shahram Rahimi. 2025. Navigating MLOps: Insights into Maturity, Lifecycle, Tools, and Careers.arXiv preprint arXiv:2503.15577(2025)

    Pith/arXiv arXiv 2025

  39. [39]

    Tom Viering and Marco Loog. 2023. The Shape of Learning Curves: A Review.IEEE Trans. Pattern Anal. Mach. Intell. 45, 6 (June 2023), 7799–7819. doi:10.1109/TPAMI.2022.3220744

    arXiv 2023

  40. [40]

    Shaoqi Wang, Aidi Pi, and Xiaobo Zhou. 2021. Elastic parameter server: Accelerating ML training with scalable resource scheduling.IEEE Transactions on Parallel and Distributed Systems33 (2021), 1128–1143

    2021

  41. [41]

    Menglu Yu, Ye Tian, Bo Ji, Chuan Wu, Hridesh Rajan, and Jia Liu. 2022. Gadget: Online resource optimization for scheduling ring-all-reduce learning jobs. InIEEE INFOCOM 2022-IEEE Conference on Computer Communications. IEEE, 1569–1578

    2022

  42. [42]

    X. Zhou, L. Chen, and T. Wu. 2024. TinyLLaVA: A Lightweight Vision–Language Assistant. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 9876–9885

    2024

  43. [43]

    Indr˙e Žliobait˙e, Marcin Budka, and Frederic Stahl. 2015. Towards cost-sensitive adaptation: When is it worth updating your predictive model?Neurocomputing150 (2015), 240–249. , Vol. 1, No. 1, Article . Publication date: December 2025. Optimal Resource Allocation for ML Model Training and Deployment under Concept Drift 19 A Technical Appendices A.1 Proof...

    2015