REVIEW 3 major objections 7 minor 19 references
Prior-aware and Context-guided Group Sampling for Active Probabilistic Subsampling
T0 review · 3 major / 7 minor · reviewed 2026-07-09 · glm-5.2
Pith's one-line read Group sampling plus dataset priors beats one-at-a-time active subsampling
desk verdict Solid empirical improvement to active subsampling; Theorem 1 doesn't match the actual architecture read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
Gumbel-Softmax reparameterization for differentiable discrete sampling; Gumbel top-k trick for selecting multiple samples simultaneously; a deterministic (fixed) learned sampling mask derived from training-data priors; an LSTM-based sampling network that encodes contextual information from previously selected samples; and a Lipschitz-constant comparison between the compositional loss of top-1 sampling and the single-function loss of top-k sampling.
What would settle it
If one could show that the one-at-a-time active sampling pipeline does not actually compose k separate task models with independently large Lipschitz constants (for instance, because parameters are shared across iterations or the effective composition is closer to a single function), then the product-of-Lipschitz-constants bound in Theorem 1 would not apply, and the theoretical justification for group sampling's smoother loss landscape would collapse. The empirical gains would then need an alternative explanation.
Extended reading notes
Core claim
The paper identifies a structural reason why selecting samples one at a time in active subsampling is harder to optimize than selecting them in groups: under the assumption that each iteration's task model has Lipschitz constant at least one, the one-at-a-time approach composes k models whose effective Lipschitz constant is the product of k constants, while the group approach uses a single model whose Lipschitz constant is just one term. Because neural network task models typically have Lipschitz constants well above one, this product grows rapidly with the number of iterations, creating a rougher optimization landscape. By combining this group-selection insight with a deterministic prior-in
Load-bearing premise
The theoretical advantage of group sampling over one-at-a-time sampling rests on modeling the one-at-a-time pipeline as a composition of k independent task models whose Lipschitz constants multiply. If the same network with shared parameters is reused across iterations, this multiplicative bound may not hold and the theoretical gap could dissolve.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes PGA-DPS, an enhancement of Active Deep Probabilistic Subsampling (A-DPS) that combines (i) a deterministic prior-informed sampling pattern derived from training data with (ii) group-based (top-k) active sampling instead of top-1. The central claims are that group sampling yields a smaller effective Lipschitz constant of the loss (Theorem 1), leading to smoother optimization, and that the combination of prior-aware fixed sampling with context-guided active group sampling consistently outperforms A-DPS, DPS, and other baselines. The method is evaluated on MNIST and CIFAR-10 classification, fastMRI knee reconstruction, and AeroRIT hyperspectral segmentation, with ablation studies on the prior/active sampling split and temperature.
Significance. The empirical contribution is substantial: four datasets, three task types, multiple independent baselines (LOUPE, VDS, greedy, RL-based methods), and consistent improvements across all settings. The provision of code, ablation studies on hyperparameters (Tables 3, 7, 8), and a temperature sensitivity analysis (Tables 9, 10) strengthen reproducibility. The practical idea of blending a fixed prior mask with a reduced number of active group-sampling iterations is simple, well-motivated, and shown to be effective, particularly in low-measurement regimes. However, the theoretical contribution (Theorem 1) does not accurately model the implemented A-DPS architecture, as detailed below.
major comments (3)
- §3.3, Theorem 1 (and Appendix A): The theorem models the DPS-top-1 loss as a composition f = f_k(f_{k-1}(...f_1(x_1)...)), yielding a product bound ∏L_r on the Lipschitz constant. However, the actual A-DPS architecture (Eqs. 7–8) does not compose task models in this manner. At iteration j, the task model f_θ is applied to the partially sampled input A_{j-1}x to produce context t_j = f_θ(A_{j-1}x), which feeds the sampling network g_j. The loss at each iteration is computed independently against the task target, and 'losses are accumulated over all iterations' (§3.2). There is no chain f_k∘f_{k-1}∘...∘f_1. The gradient path from early samples to later losses does pass through alternating task model and sampling network evaluations, but the product bound ∏L_r captures only task model Lipschitz constants and omits the sampling network g_j entirely. Thus the claimed inequality L_k ≤ ∏L_r and
- the conclusion that 'DPS-top-k exhibits a smaller effective Lipschitz constant than DPS-top-1' do not follow from the actual architecture. The empirical gains (Tables 1–6) are consistent and do not depend on this theorem, but the theoretical contribution as stated is not valid. The authors should either (a) revise Theorem 1 to model the actual A-DPS loss accumulation structure (sum of independent per-iteration losses with gradient backpropagation through shared parameters) and re-derive the Lipschitz bound accordingly, or (b) reframe the theorem as a motivating heuristic rather than a formal result, making clear that it describes an idealized composition not present in the implementation.
- §3.3 and Table 3: The optimal (P_s, A_s) configuration varies across tasks: (60%, 20%) for MNIST, (10%, 20%) for CIFAR-10, (30%, 30%) for MRI, and (80%, 20%) for HSI segmentation. The paper provides a heuristic recommendation in §6 (use the DPS vs. A-DPS performance gap as an indicator of Lipschitz characteristics), but this is not validated quantitatively. Given that P_s ranges from 10% to 80% across tasks, the practical guidance for selecting these hyperparameters remains underspecified. A sensitivity analysis showing performance degradation as a function of deviation from the optimal P_s (beyond the partial results in Table 7) would strengthen the claim that the method is robust to hyperparameter choice.
minor comments (7)
- §3.2, Eq. (8): The notation ϕ_j = g_j(t_j) uses j as both the iteration index and a subscript for the sampling network g_j, but it is unclear whether g_j denotes a distinct network per iteration or the same network with different inputs. Clarification needed.
- Table 3: The entry for P_s=70%, A_s=40% is marked '—' with no explanation. Please state why this configuration was not evaluated.
- Figure 1 caption: 'DPS picks 31 samples in one step, A-DPS over 31 iterations, and PGA-DPS in just 3 iterations' — the number 31 appears task-specific (MNIST at ~4% sampling) but is not defined in the caption context. Consider clarifying.
- §4.1.2: The sentence 'the proportions of prior (deterministic) sampling and active sampling are fixed to 60 and 20 %' is missing a unit or context — should read '60% and 20% of the target samples' for consistency with later sections.
- Appendix B.4: 'tow downsampling operations' should be 'two downsampling operations'.
- Table 2: A-DPS shows a performance decrease beyond r=14% (e.g., 70.4 at r=10% vs. 68.3 at r=20%), which the authors attribute to 'an inflated Lipschitz constant.' This explanation relies on Theorem 1, which as noted above does not accurately model the architecture. An alternative empirical explanation (e.g., overfitting of the sampling network at higher sampling ratios) should be discussed.
- The paper states (§3.2) that 'the DPS-top-k approach showed improved performance' over DPS-top-1, citing Huijben et al. (2020a). Since this is a central motivation, a brief explanation of why top-k outperforms top-1 in the original DPS work would help readers unfamiliar with that result.
Circularity Check
No circularity: Theorem 1 is a standard Lipschitz composition bound; empirical results validated against external baselines
full rationale
The paper's theoretical contribution (Theorem 1) is derived from standard properties of Lipschitz composition: if f = f_k ∘ f_{k-1} ∘ ... ∘ f_1, then the Lipschitz constant of the composition is the product ∏L_r, while a single function f_k has constant L_k. This is a textbook result, not a self-citation chain or a fitted-then-predicted quantity. The empirical results are validated against multiple external baselines (DPS, A-DPS, LOUPE, VDS, RL methods) from independent research groups, and the hyperparameters (Ps, As) are tuned via ablation, not forced by construction. The A-DPS baseline (Van Gorp et al., 2021) is cited as prior work by different authors. While the skeptic raises a valid correctness concern—namely that Theorem 1 models the DPS-top-1 loss as a composition f = f_k∘...∘f_1 when the actual A-DPS architecture (Eqs. 7–8) applies f_θ independently at each iteration without composing task models—this is a modeling mismatch issue, not circularity. The theorem does not define its inputs in terms of its outputs, does not fit a parameter and then call the fit a prediction, and does not rely on a self-citation chain to establish its core inequality. The derivation is self-contained as a mathematical statement, even if its applicability to the actual architecture is debatable. No step in the paper's claimed derivation chain reduces to its inputs by construction.
Assumptions & free parameters
free parameters (4)
- Ps (prior sampling ratio) =
task-dependent: 60% (MNIST), 10% (CIFAR-10), 30% (MRI), 80% (HSI)
- As (active sampling ratio) =
task-dependent: 20% (MNIST, HSI), 20% (CIFAR-10), 30% (MRI)
- tau (temperature) =
2
- phi_1 (prior logits) =
learned (size 784/1024/208/51 depending on task)
assumptions (3)
- domain assumption Neural networks typically have Lipschitz constants much greater than one, except in the trivial near-identity case.
- ad hoc to paper DPS-top-1 involves a composition of k separate task functions f = f_k(f_{k-1}(...f_1(x_1)...)), while DPS-top-k involves a single function f_k over the group.
- standard math The Gumbel-softmax trick provides a differentiable approximation to categorical sampling suitable for backpropagation.
Cite this review
Pith. "Pith review of Prior-aware and Context-guided Group Sampling for Active Probabilistic Subsampling." pith.science (2026). https://pith.science/paper/KGYVXSIY
@misc{pith2026260707083,
author = {Pith},
title = {Pith review of: Prior-aware and Context-guided Group Sampling for Active Probabilistic Subsampling},
year = {2026},
howpublished = {\url{https://pith.science/paper/KGYVXSIY}},
note = {Machine review of arXiv:2607.07083}
}
read the original abstract
Subsampling significantly reduces the number of measurements, thereby streamlining data processing and transfer overhead, and shortening acquisition time across diverse real-world applications. The recently introduced Active Deep Probabilistic Subsampling (A-DPS) approach jointly optimizes both the subsampling pattern and the downstream task model, enabling instance- and subject-specific sampling trajectories and effective adaptation to new data at inference time. However, this approach does not fully leverage valuable dataset priors and relies on top-1 sampling, which can impede the optimization process. Herein, we enhance A-DPS by integrating a deterministic (fixed) prior-informed sampling pattern derived from the training dataset, along with group-based sampling via top-k sampling, to achieve more robust optimization, method we call Prior-aware and context-guided Group-based Active DPS (PGA-DPS). We also provide a theoretical analysis supporting improved optimization via group sampling, and validate this with empirical results. We evaluated PGA-DPS on three tasks: classification, image reconstruction, and segmentation, using the MNIST, CIFAR-10, fastMRI knee, and hyperspectral AeroRIT datasets, respectively. In every case, PGA-DPS outperformed A-DPS, DPS, and all other sampling methods.
Figures
Figures from the paper (1 more)
Reference graph
Works this paper leans on
-
[1]
Peter L Bartlett, Steven N Evans, and Philip M Long. Representing smooth functions as composi- tions of near-identity functions with implications for deep network optimization.arXiv preprint arXiv:1804.05012,
-
[2]
11 Published as a conference paper at ICLR 2026 Guang-Hong Chen, Jie Tang, and Shuai Leng. Prior image constrained compressed sensing (piccs): a method to accurately reconstruct dynamic ct images from highly undersampled projection data sets.Medical physics, 35(2):660–663,
work page 2026
-
[3]
Deep Residual Learning for Compressed Sensing CT Reconstruction via Persistent Homology Analysis
Yo Seob Han, Jaejun Yoo, and Jong Chul Ye. Deep residual learning for compressed sensing ct reconstruction via persistent homology analysis.arXiv preprint arXiv:1611.06391,
-
[4]
Deep probabilistic subsampling for task-adaptive compressed sensing
Iris Huijben, Bastiaan S Veeling, and Ruud JG van Sloun. Deep probabilistic subsampling for task-adaptive compressed sensing. In8th International Conference on Learning Representations, ICLR 2020, 2020a. Iris AM Huijben, Bastiaan S Veeling, Kees Janse, Massimo Mischi, and Ruud JG van Sloun. Learn- ing sub-sampling and signal recovery with applications in ...
work page 2020
-
[5]
Self-Supervised Deep Active Accelerated MRI
Kyong Hwan Jin, Michael Unser, and Kwang Moo Yi. Self-supervised deep active accelerated mri. arXiv preprint arXiv:1901.04547,
work page Pith review arXiv 1901
-
[6]
Adam: A Method for Stochastic Optimization
Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization.arXiv preprint arXiv:1412.6980,
-
[7]
Photo-realistic sin- gle image super-resolution using a generative adversarial network
12 Published as a conference paper at ICLR 2026 Christian Ledig, Lucas Theis, Ferenc Husz ´ar, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photo-realistic sin- gle image super-resolution using a generative adversarial network. InProceedings of the IEEE conference on computer vision ...
work page 2026
-
[8]
Learning smooth neural functions via lipschitz regularization
Hsueh-Ti Derek Liu, Francis Williams, Alec Jacobson, Sanja Fidler, and Or Litany. Learning smooth neural functions via lipschitz regularization. InACM SIGGRAPH 2022 Conference Proceedings, pp. 1–13,
work page 2022
Show all 19 references
-
[9]
Scalable learning-based sampling optimization for compressive dynamic mri
Thomas Sanchez, Baran G ¨ozc¨u, Ruud B van Heeswijk, Armin Eftekhari, Efe Ilıcak, Tolga C ¸ ukur, and V olkan Cevher. Scalable learning-based sampling optimization for compressive dynamic mri. InICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Pro...
2020
-
[10]
Learning to scan: A deep reinforcement learning approach for personalized scanning in ct imaging.arXiv preprint arXiv:2006.02420,
13 Published as a conference paper at ICLR 2026 Ziju Shen, Yufei Wang, Dufan Wu, Xu Yang, and Bin Dong. Learning to scan: A deep reinforcement learning approach for personalized scanning in ct imaging.arXiv preprint arXiv:2006.02420,
2026 arXiv
-
[11]
Joint learning of cartesian under sampling andre construction for accelerated mri
Tomer Weiss, Sanketh Vedula, Ortal Senouf, Oleg Michailovich, Michael Zibulevsky, and Alex Bronstein. Joint learning of cartesian under sampling andre construction for accelerated mri. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing...
2020
-
[12]
fastmri: An open dataset and benchmarks for accelerated mri.arXiv preprint arXiv:1811.08839,
Jure Zbontar, Florian Knoll, Anuroop Sriram, Tullie Murrell, Zhengnan Huang, Matthew J Muckley, Aaron Defazio, Ruben Stern, Patricia Johnson, Mary Bruno, et al. fastmri: An open dataset and benchmarks for accelerated mri.arXiv preprint arXiv:1811.08839,
-
[13]
Extending loupe for k-space under-sampling pattern optimization in multi- coil mri
Jinwei Zhang, Hang Zhang, Alan Wang, Qihao Zhang, Mert Sabuncu, Pascal Spincemaille, Thanh D Nguyen, and Yi Wang. Extending loupe for k-space under-sampling pattern optimization in multi- coil mri. InMachine Learning for Medical Image Reconstruction: Third International Worksh...
2020
-
[14]
Reducing uncertainty in undersampled mri reconstruction with active acquisition
Zizhao Zhang, Adriana Romero, Matthew J Muckley, Pascal Vincent, Lin Yang, and Michal Drozdzal. Reducing uncertainty in undersampled mri reconstruction with active acquisition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 2049– 2058,
-
[15]
We provide a proof for the effect of DPS-top-1 sampling on the Lipschitzness of the loss
14 Published as a conference paper at ICLR 2026 APPENDIX A PROOF In this Appendix, we prove the Theorem 1 in Method. We provide a proof for the effect of DPS-top-1 sampling on the Lipschitzness of the loss. Theorem 1.Letf 1, f2, . . . , fk be task models atj th iteration, wher...
2026
-
[16]
Each layer is followed by batch normalization, a ReLU activation function, and a 2×2 max pooling operation
B.2 CIFAR-10CLASSIFICATION The classification networkf θ(·)consists of four convolutional layers with channel sizes of 32, 64, and 128, respectively, and a kernel size of 3×3. Each layer is followed by batch normalization, a ReLU activation function, and a 2×2 max pooling oper...
-
[17]
Each layer is followed by a ReLU activation function except for the final layer
B.3 MRIRECONSTRUCTION The proximal neural networkP ψ consists of four convolutional layers with channel size of 16,16,16, and 1, respectively, and a kernel size of 3×3. Each layer is followed by a ReLU activation function except for the final layer. For the image prior regular...
2017
-
[18]
16 Published as a conference paper at ICLR 2026 B.4 HYPERSPECTRAL IMAGE SEGMENTATION For the segmentation networkf θ(·), we use a residual U-Net consisting of 6 ResNet blocks, as pro- posed in the pix2pix (Isola et al., 2017; Zhu et al., 2017). The input channel is set to 51, ...
2026
-
[19]
Table 7 shows that PA-DPS outperforms the baselines in almost all configurations, although the optimal configuration shifts slightly across sampling ratios
C RESULTS C.1 MNISTCLASSIFICATION We evaluated multipleP svalues under anAsof 20% and multipleAsvalues under aP sof 60%. Table 7 shows that PA-DPS outperforms the baselines in almost all configurations, although the optimal configuration shifts slightly across sampling ratios....
2026
Reviewed July 9, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.