{"id":"4294abf8-6eb4-43f9-9c85-bd7d4831e22c","arxiv_id":"1908.05764","paper_version":5,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"DPS jointly learns a probabilistic sub-sampling pattern and a neural reconstruction network, producing task-specific sparse sampling schemes for partial Fourier and ultrasound data.","lead":"This paper introduces a deep learning framework, Deep Probabilistic Sub-sampling (DPS), that learns which parts of a signal to keep before compressing and then learns to recover the information needed for a given imaging task. It is relevant because it shows how to design practical, task-specific sampling patterns for bandwidth-limited ultrasound imaging.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Most load-bearing concern: the central 'fixed, implementable pattern' claim rests on a single stochastic realization, and Appendix A shows that for channel-sub-sampling Doppler the trained distributions are not near one-hot; reported gains may be realization-specific.","rationale":"The reader's verdict (CONDITIONAL) is appropriate. I reviewed the central claim: DPS jointly learns a task-driven sub-sampling pattern and a recovery network, and the learned pattern is fixed and hardware-implementable. For this claim to hold, the training procedure must yield a pattern that is representative of the learned distribution. The paper's own Appendix A shows a counterexample: channel sub-sampling for Doppler did not converge to near-one-hot distributions. Nevertheless, results for that configuration are reported from a single stochastic realization. This makes the quantitative results (Fig. 8b) and the implementability claim fragile. This is not a matter of theoretical soundness of Gumbel-Softmax; it is an empirical robustness gap that can be closed by repeated sampling. I therefore agree with the reader's weakest_assumption concerning stochastic realizations, and I highlight it as the most load-bearing issue. The surrogate-gradient bias is real but secondary: since the paper makes no formal optimality claim, the key question is whether the final sampled pattern is good in practice, and that question is currently unanswered for the non-converged channel-Doppler case. My recommendation remains CONDITIONAL: the paper is a solid contribution but needs realization-level error bars (or a deterministic mask extraction procedure) before the central claim is fully supported.","tokens_in":19286,"tokens_out":5878,"duration_ms":58278,"concrete_test":"Using the trained Φ from Section V-C Doppler channel sub-sampling (Appendix A, Fig. 10d) and the trained task model gθ, draw K=50 independent realizations of the sampling mask via Eq. (7) with masking without replacement (Algorithm 1). For each realization, compute the Doppler MSE over the full hold-out test set, and report the mean and standard deviation. If the standard deviation is non-negligible compared to the DPS-vs-uniform gap in Fig. 8b (or if the uniform baseline beats the mean DPS result), the reported single-realization result is not representative and the fixed-pattern claim fails for this task. As a secondary check, do the same for slow-time Doppler (Section V-B) and for the Fourier experiment; if all variances are small, the realization concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that DPS yields a fixed, task-driven sub-sampling pattern that is straightforwardly implementable (Section III-B, III-D). The training procedure only learns categorical distributions over the N candidates (Eq. 4-5); the reported pattern is one random draw from those distributions. For most US tasks the distributions converged to near-one-hot, but Appendix A explicitly states that for channel sub-sampling with color Doppler this was not the case (Fig. 10d). Section V-C and Fig. 8b nevertheless report DPS performance for that task using a single realization. Since the trained logits still assign substantial probability to many channels, a different realization could yield a different sparse array with markedly different grating-lobe behaviour and Doppler MSE. Without repeated realizations or error bars, the quantitative advantage of DPS over uniform sampling for channel Doppler is not established, and the claim that the learned pattern can be fixed for hardware implementation is unsupported for exactly the application where the distributions are least concentrated. The Gumbel-Softmax surrogate (Eq. 8) is a secondary issue: even if the surrogate is biased, the empirical question is whether the final sampled pattern is typically good; the paper does not provide the statistics needed to answer that question. The single-realization/non-convergence gap is therefore the more immediate threat to the central claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Deep Probabilistic Sub-sampling (DPS), a framework that jointly learns a categorical distribution over candidate measurement indices and a neural task model in an end-to-end fashion. Sampling is performed without replacement via a Gumbel-max trick with masking, and gradients are obtained through a straight-through Gumbel-Softmax surrogate with temperature annealing. The method is validated on three settings: sparse signal recovery from partial Fourier measurements, slow-time sub-sampling for ultrasound B-mode and color Doppler imaging, and channel sub-sampling for sparse array design. The authors report that DPS yields task-specific patterns, that learned patterns are comparable or superior to uniform and random fixed sampling when combined with the learned task model, and that the resulting fixed patterns are hardware-implementable.","tokens_in":19600,"tokens_out":8017,"duration_ms":88774,"significance":"If the reported results are robust, the work is significant: it offers a practical way to learn hard, task-driven sub-sampling masks with a trainable downstream network, avoiding the hardware challenges of randomly weighted linear measurement schemes. The method is clearly described with pseudocode and the masking mechanism for sampling without replacement is coherent. The use of in-vivo porcine ultrasound data and the comparison with a hand-designed full-sum-coarray sparse array are strong points, as is the interpretable finding that B-mode and Doppler tasks induce qualitatively different learned sampling patterns. The main limitations are the reliance on single stochastic realizations for some reported patterns and the absence of error bars or repeated-run statistics in the quantitative comparisons.","major_comments":[{"comment":"The central claim that DPS produces a fixed, hardware-implementable pattern that outperforms uniform sampling is not established for channel sub-sampling with color Doppler. Appendix A explicitly reports that the learned probability distributions for this task did not converge to near-one-hot (Fig. 10d), yet the quantitative DPS results in Fig. 8b and the qualitative images in Fig. 7 are based on a single stochastic realization of those distributions. Different realizations can yield different sparse arrays with different grating-lobe behavior and different Doppler MSE, so the reported advantage over uniform sampling may be realization-specific. Please report the mean and spread of test MSE over multiple draws from the trained distributions, and specify the procedure by which a single deployment pattern would be selected from a non-concentrated distribution.","section":"Section V-C / Fig. 8 / Appendix A"},{"comment":"The entropy penalty LS is computed from the unmasked probabilities πm defined in Eq. (5), while the actual forward sampling in Eq. (7) uses masked logits wm−1 + φm to enforce sampling without replacement. Penalizing the unmasked entropy does not guarantee that the conditional distribution used at row m is concentrated after earlier rows have removed candidates. This mismatch plausibly contributes to the non-convergence reported in Appendix A for channel Doppler and should be fixed or analyzed; otherwise the training objective does not directly promote the one-hot patterns on which the 'fixed pattern' claim rests.","section":"Section III-D, Eq. (11)"},{"comment":"All quantitative comparisons are reported as point estimates without error bars, confidence intervals, or significance tests. Because DPS training is stochastic (Gumbel noise, mini-batch sampling) and the final evaluation itself uses a random realization, the observed differences between DPS and Random+LISTA or between DPS and Uniform+task-model could be within run-to-run variability. Please report results over at least three independent training runs and multiple test realizations for the main comparisons.","section":"Figs. 4, 6, 8"}],"minor_comments":[{"comment":"The transducer system is referred to as 'Verasonics Vantrage'; this appears to be a typo for 'Verasonics Vantage'.","section":"Section IV-B1"},{"comment":"The caption contains a duplicated MSE value ('MSE: 2.1e-62.1e-6') and the typo 'task model.s'; both should be corrected.","section":"Fig. 9 caption"},{"comment":"The notation ∇φm am := ∇φm E[softmaxτ(...)] is nonstandard because the left-hand side denotes a gradient of a hard one-hot sample. Please clarify that this is the surrogate gradient used in backpropagation, not the true gradient of am.","section":"Section III-B, Eq. (8)"},{"comment":"The statement that the learned sensing matrix 'showed to be RIP-compliant' should specify how the check was performed, for example whether all M×K submatrices were tested and what numerical criterion was used.","section":"Section V-A"}],"recommendation":"major_revision","confidential_remarks":"The main barrier is the single-realization and single-run evaluation for the channel-Doppler experiment, combined with the entropy-penalty mismatch. These are fixable with additional experiments rather than being fundamental flaws. The authors cite their own ICLR paper [28] for the DPS method; the editor may wish to verify that the journal version's novelty relative to that prior publication is clearly delineated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about arXiv:1908.05764. First, the core idea is real and was genuinely new at the time: instead of learning weighted linear measurements like most learned-CS work, DPS learns categorical distributions over which signal elements to keep, using Gumbel-Softmax plus a masking trick for sampling without replacement. That closes a practical gap—weighted combinations are hard to build into hardware, discrete sub-sampling masks are not. Second, the paper is honest in ways that make it better than average. It reports an essentially negative result for slow-time B-mode (the learned pattern is nearly uniform, performance matches uniform sampling), and Appendix A states openly that for channel sub-sampling with color Doppler the trained distributions were far from one-hot.\n\nThe math in Section III is coherent. The row-wise one-hot constraint on the sampling matrix, the w_{m-1} masking for sampling without replacement, the straight-through Gumbel gradient with temperature annealing, and the entropy penalty are all set up correctly, and the pseudocode is clear enough to implement. The ultrasound targets (B-mode envelope, Kasai Doppler) are computed from fully sampled data, so there is no circularity. Comparing the learned sparse array against a fixed full-sum-coarray design is the right kind of check. The Fourier experiment is a useful sanity check: DPS matches random sampling, and both beat uniform, which is what CS theory would suggest.\n\nThe soft spot is the one the authors themselves flag. The central claim that DPS yields a fixed, hardware-implementable pattern is not supported for exactly the channel-Doppler case where the distributions are least concentrated. The reported result is a single stochastic draw from distributions that still assign substantial probability across many channels, and a different draw could give a different sparse array with different grating-lobe behavior. There are no error bars or repeated-realization statistics anywhere in the results, and the ultrasound validation uses one train/test split from a two-animal cohort. That limits the strength of the channel-Doppler comparison in Fig. 8b, though the Fourier and slow-time results are more robust. The Gumbel-Softmax surrogate being heuristic is a common field-wide issue, not a paper-specific flaw. Minor point: the method overlaps with the authors' own ICLR 2020 paper, and the journal version should state clearly what is added beyond it (the ultrasound applications).\n\nThis paper is for anyone working on learned sensing, compressed ultrasound, or hardware-aware acquisition. I would send it to peer review and would cite the DPS method. The main things I'd ask a revision to add: repeated realizations and error bars for the non-converged distributions, and a concrete statement of how the pattern is selected at deployment time.","headline":"A genuinely useful and honestly reported learned-subsampling paper; the one real gap is that the 'fixed pattern' claim rests on a single draw from distributions that Appendix A shows never converged for channel Doppler.","tokens_in":20147,"tokens_out":5158,"would_cite":true,"duration_ms":49386,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A single end-to-end training procedure, Deep Probabilistic Sub-sampling, learns fixed sub-sampling masks that retain task-relevant information at least as well as random sampling and better than uniform sampling for ultrasound B-mode and…","keywords":["deep probabilistic sub-sampling","task-adaptive sampling","compressed sensing","Gumbel-Softmax","ultrasound imaging","Doppler imaging","sparse arrays","learned reconstruction"],"falsifier":"Train DPS multiple times from different random seeds on the same in-vivo ultrasound data, fix one mask per run, and evaluate each on the hold-out test set; if the test MSEs across realizations vary widely, or if a fixed random mask at the same sub-sampling factor matches or beats the learned mask, then the claim that learned patterns retain task-relevant information better than standard sampling is not supported.","tokens_in":19101,"feed_emoji":"🩺","tokens_out":8514,"duration_ms":72998,"temperature":0.7,"pith_summary":"This paper tries to establish that the sampling pattern itself, which elements of a signal get measured, can be learned from data at the same time as the network that turns those measurements into a useful output. The proposed method, Deep Probabilistic Sub-sampling (DPS), models each of the $M$ measurements as a categorical draw over the $N$ candidate sample locations, uses the Gumbel-Softmax relaxation to backpropagate through the discrete choice, and after training fixes a mask drawn from the learned distributions. The authors show that DPS produces task-specific patterns: near-uniform for B-mode ultrasound, clustered slow-time ensembles for Doppler, and center-weighted channel selection for sparse arrays. The payoff, if correct, is that imaging systems can cut data rates and power consumption using fixed, hardware-friendly non-uniform sampling that retains task-relevant information at least as well as random or hand-designed sampling.","feed_headline":"Learned sampling beats fixed patterns in ultrasound tasks","feed_subtitle":"A joint training loop produces fixed, hardware-ready masks that preserve what B-mode and Doppler imaging need.","key_machinery":"The load-bearing object is the DPS generative sampling model: for each measurement $m \\in \\{1,\\dots,M\\}$, a categorical random variable $r_m \\sim \\text{Cat}(N,\\pi_m)$ over the $N$ candidate positions, with probabilities $\\pi_{m,n} = \\exp(\\phi_{m,n}) / \\sum_i \\exp(\\phi_{m,i})$ derived from trainable logits $\\phi_{m,n}$. A one-hot sample is drawn through the Gumbel-max trick, and sampling without replacement is enforced by adding $-\\infty$ to previously chosen positions via a mask $w_{m-1}$. Training uses the straight-through Gumbel estimator: the forward pass commits to the hard $\\arg\\max$ sample, while the gradient is computed through $\\mathrm{softmax}_\\tau(w_{m-1} + \\phi_m + e_m)$, with temperature annealed from $5.0$ to $0.5$ and an entropy penalty pushing the distributions toward one-hot masks. This machinery is what makes joint, differentiable optimization of a discrete sampling pattern and a downstream network possible.","core_discovery":"The paper's central claim is that a task-driven sub-sampling mask can be optimized end-to-end without relaxing the constraint that measurements are direct selections of input elements. DPS represents the mask as a matrix of trainable logits $\\Phi \\in R^{M \\times N}$, draws rows as one-hot samples without replacement via the Gumbel-max trick, and trains $\\Phi$ jointly with a neural task model $g_\\theta$ by minimizing mean squared error plus an entropy penalty. Because the forward pass uses hard one-hot sampling while the backward pass uses the softmax temperature surrogate, the learned pattern converges to a fixed, discrete mask. Across partial Fourier recovery, slow-time ultrasound, and channel sub-sampling, the resulting masks match or outperform fixed uniform and random sampling when paired with the jointly trained network, and in Doppler imaging the learned ensemble pattern substantially reduces aliasing compared with naive interpolation plus the Kasai estimator.","pith_inferences":["Beyond the paper's claims, the same sub-sampling formalism would transfer naturally to MRI k-space or sparse-view CT, since those modalities already acquire partial Fourier or projection measurements; the partial-Fourier experiment is the closest evidence.","Because the reported masks are single stochastic realizations of trained distributions, a seed-variability study would tell whether the learned pattern itself is stable; the paper does not report this, so the realized mask should be read as one sample, not as the unique optimum.","One could test the method's sensitivity to the straight-through surrogate by training with several temperature schedules or entropy penalty weights; if performance is strongly schedule-dependent, the relaxation's behavior, not the sampling concept, may be the limiting factor."],"forward_implications":["Once training is done, the learned mask is a fixed set of sample locations, so it can be implemented by non-uniform analog-to-digital conversion, sparse array design, or slow-time pulsing schemes without any per-acquisition optimization.","The same training procedure adapts the pattern to the task: B-mode recovery drives near-uniform slow-time sampling, Doppler recovery drives ensemble-like clustered sampling, and channel sub-sampling drives center-weighted arrays.","Jointly trained sampling plus task model outperforms fixed sampling with a trained model, and for Doppler it avoids the strong aliasing seen when interpolating frames before applying the Kasai auto-correlator.","The unfolded learned-ISTA task model recovers sparse signals from partial Fourier measurements faster than 300-iteration ISTA by a factor over 1000, with better MSE at high sub-sampling factors.","Learned sparse arrays can match a hand-designed full-sum-coarray array in B-mode MSE, suggesting DPS discovers known good array geometries from data."],"supporting_citations":[{"why":"Supplies the straight-through Gumbel estimator used to backpropagate through the discrete sub-sampling samples.","marker":"[26]"},{"why":"Provides the continuous relaxation of discrete random variables (concrete distribution) underlying the softmax surrogate gradient.","marker":"[27]"},{"why":"Establishes random partial Fourier sampling as the RIP-compliant baseline that DPS must match or beat.","marker":"[19]"},{"why":"Introduces the unrolled LISTA architecture that the task model for sparse-signal recovery is based on.","marker":"[46]"},{"why":"Supports the claim that few-unfolding LISTA can outperform iterative ISTA, motivating the fast task model.","marker":"[47]"},{"why":"Provides the full-sum-coarray sparse-array design used as the hand-designed baseline for channel sub-sampling.","marker":"[35]"},{"why":"Defines the Kasai auto-correlator used both to generate Doppler targets and as the non-learned Doppler estimator in comparisons.","marker":"[51]"}],"fun_headline_variants":["Deep learning finds optimal ultrasound sampling patterns","AI designs sampling masks that beat fixed patterns","Learned masks improve ultrasound reconstruction","Task-driven sub-sampling outperforms uniform sampling"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim rests on the assumption that the straight-through Gumbel-Softmax surrogate gradient, with the chosen temperature annealing and entropy penalty, actually steers the hard sampling mask toward a near-optimal pattern; this is supported empirically but not by any formal guarantee, and the evaluated masks are from a single stochastic realization.","fun_headline_variants_meta":{"raw":{"variants":["Deep learning finds optimal ultrasound sampling patterns","AI designs sampling masks that beat fixed patterns","Learned masks improve ultrasound reconstruction","Task-driven sub-sampling outperforms uniform sampling"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000175,"raw_usage":{"total_tokens":1297,"prompt_tokens":968,"completion_tokens":329,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":584,"completion_tokens_details":{"reasoning_tokens":276}},"tokens_in":584,"tokens_out":329,"duration_ms":3927,"temperature":1.0,"reasoning_tokens":276,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:04:19.557496+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train DPS multiple times from different random seeds on the same in-vivo ultrasound data, fix one mask per run, and evaluate each on the hold-out test set; if the test MSEs across realizations vary widely, or if a fixed random mask at the same sub-sampling factor matches or beats the learned mask, then the claim that learned patterns retain task-relevant information better than standard sampling is not supported.","supporting_citations":[{"cited_title":"Categorical reparametrization with gumbel-softmax,","cited_arxiv_id":null,"evidence_quote":"Supplies the straight-through Gumbel estimator used to backpropagate through the discrete sub-sampling samples."},{"cited_title":"Compressive sampling,","cited_arxiv_id":null,"evidence_quote":"Establishes random partial Fourier sampling as the RIP-compliant baseline that DPS must match or beat."},{"cited_title":"Learning fast approximations of sparse coding,","cited_arxiv_id":null,"evidence_quote":"Introduces the unrolled LISTA architecture that the task model for sparse-signal recovery is based on."},{"cited_title":"Theoretical linear convergence of unfolded ista and its practical weights and thresholds,","cited_arxiv_id":null,"evidence_quote":"Supports the claim that few-unfolding LISTA can outperform iterative ISTA, motivating the fast task model."},{"cited_title":"Sparse convolutional beamforming for ultrasound imaging,","cited_arxiv_id":null,"evidence_quote":"Provides the full-sum-coarray sparse-array design used as the hand-designed baseline for channel sub-sampling."},{"cited_title":"Real-time two-dimensional blood flow imaging using an autocorrelation technique,","cited_arxiv_id":null,"evidence_quote":"Defines the Kasai auto-correlator used both to generate Doppler targets and as the non-learned Doppler estimator in comparisons."}],"review_version":1}