{"id":"25ba4313-4f2a-48d4-950e-eb679dbdb3e7","arxiv_id":"1908.02144","paper_version":4,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Batch active learning is reformulated as sparse subset approximation of the expected complete-data log-posterior, solved with Frank-Wolfe and random projections.","lead":"An active learning method that selects entire batches of data points to label, by treating batch choice as a sparse approximation to the full model posterior. It scales to large datasets and produces diverse queries, and is shown to be competitive on regression and image classification.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section 3 asserts without proof that approximating the expected complete-data log posterior yields a close posterior; the missing transfer bound is the weakest load-bearing link.","rationale":"The reader identified the missing transfer guarantee as the weakest assumption, and my reading agrees. This is the single most load-bearing concern because the paper's central claim has two components: an empirical component (ACS-FW is scalable and competitive) and a theoretical component (batch AL is principled as sparse subset approximation). The empirical component is supported by the experiments, though not by formal guarantees. The theoretical component relies entirely on the assertion in Section 3 that function-space approximation error translates to posterior closeness. That assertion is not proven, and moreover the objective targets the expectation of the log posterior rather than the log of the expected posterior, so the transfer is not a routine corollary of existing coreset bounds. A small, fully computable experiment can test whether the proxy has any empirical validity. I would not change the verdict: CONDITIONAL remains appropriate because the concern is addressable (add a transfer bound or soften the 'principled' claim) and the empirical results still provide value. Other weaknesses, such as the projection step from continuous Frank-Wolfe weights to binary batches and the use of Student-t likelihoods in experiments, are real but secondary; they would matter even if the transfer guarantee were supplied, but the missing transfer is the first hurdle.","tokens_in":15943,"tokens_out":11168,"duration_ms":127028,"concrete_test":"On a small synthetic Bayesian linear regression problem (e.g., input dimension d=2, pool size M=20, batch size b=5), enumerate all C(20,5) batches. For each batch S compute (i) the sparse-subset objective error ||L - L(1_S)||^2 using the closed-form Fisher inner product from Eq. (11), and (ii) the exact expected posterior entropy after acquiring labels, E_{Y'}[H[theta | D0 union (X', Y')]], by drawing Y' from the predictive posterior and using the closed-form Gaussian posterior update. Report the Spearman rank correlation between (i) and (ii). If batches with small objective error do not systematically achieve lower expected posterior entropy, the sparse-subset proxy is not selecting batches that optimize the stated active-learning goal.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that ACS-FW is a principled Bayesian batch AL method rests on the assertion in Section 3 that a good function-space approximation to L = sum_m L_m implies the resulting posterior is close to the expected complete-data posterior. No theorem or bound is provided connecting ||L - L(w)|| from Eq. (5) or Eq. (6) to any divergence between the batch-updated posterior p(theta | D0 union D') and the complete-data posterior p(theta | D0 union Dp), nor to the expected posterior entropy objective in Eq. (2). In the Hilbert coreset literature (Campbell and Broderick, 2019), such transfer guarantees are derived from log-likelihood geometry; here the analogous step is absent. The gap is not merely a missing technical lemma: the quantity being approximated is the expectation of the complete-data log posterior, not the log of the expected complete-data posterior, and the actual posterior after labeling uses realized labels Y', not expectations over Yp. Thus even an exact solution of the sparse-subset problem would not by itself imply that the selected batch is near-optimal for the AL objective. If this proxy is unreliable, the 'principled' motivation for ACS-FW collapses, even though the algorithm may still work empirically as a diversity-promoting heuristic.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ACS-FW, a Bayesian batch active learning method that constructs each batch by solving a sparse subset approximation to the expected complete-data log posterior. The objective is relaxed into a Hilbert-space norm minimization and solved with Frank-Wolfe, with closed-form inner products for Bayesian linear and probit regression, and a random-projection estimator of the Euclidean inner product for arbitrary tractable likelihoods. The authors compare ACS-FW against random selection, MAXENT, BALD, k-center, and k-medoids on UCI regression datasets, cifar10, SVHN, and Fashion MNIST, reporting that ACS-FW outperforms random selection and matches or exceeds existing batch methods, with favorable runtime scaling. The theoretical motivation is that a small function-space approximation error implies a close posterior, but no transfer bound is provided.","tokens_in":16179,"tokens_out":8487,"duration_ms":80628,"significance":"The paper offers a novel and elegant reformulation of batch active learning as sparse subset approximation of the expected complete-data log posterior, drawing on the Hilbert coreset toolbox. If the theoretical gap identified below can be closed, the method would provide a scalable, model-agnostic Bayesian alternative to existing batch AL heuristics. The closed-form connections to BALD and leverage scores for linear models, the random-projection extension to general likelihoods, and the empirical results on datasets with hundreds of thousands of points are the main strengths. The paper also releases source code, which supports reproducibility. However, the missing transfer guarantee between the approximation objective and the actual AL objective is a substantive caveat; the empirical gains over strong baselines such as MAXENT are real but modest on the small datasets.","major_comments":[{"comment":"The paper's central claim that ACS-FW is a principled Bayesian batch AL method rests on the assertion in Section 3 that a good function-space approximation to L implies that the resulting posterior is close to the expected complete-data posterior. No theorem or bound is provided connecting ||L - L(w)|| to a divergence between p(θ|D0∪D') and p(θ|D0∪Dp), nor to the expected posterior entropy objective in Eq. (2). Because Eq. (4) targets the expectation of the log posterior while Eq. (2) is an expectation over posterior entropies with realized labels, even an exact optimizer of Eq. (5) would not automatically be near-optimal for the AL objective. Please add a quantitative transfer result (or a counterexample analysis) or explicitly re-frame the sparse-subset objective as a heuristic surrogate.","section":"Section 3, Eqs. (4)-(6)"},{"comment":"For the weighted Euclidean inner product used in the random-projection variant, the constant entropy term H[y_m|x_m,D0] in L_m (Eq. (4)) does not vanish, since no gradient is taken. This term is irrelevant for the posterior, but it contributes to the norm and to the inner products estimated in Eq. (16), so the objective (5) may be dominated by label-independent constants. Please analyze how this affects the selected batch; if the entropy term is intended to be included, a justification is needed, and if not, it should be removed from L_m before projection.","section":"Section 5, Eqs. (9), (16)"}],"minor_comments":[{"comment":"In Section 4 (Eq. (12)) and Appendix B.1, the statement that exp(2 α_BALD) is proportional to α_ACS is incorrect: exp(2 α_BALD) = 1 + x_n^T Σ_θ x_n / σ_0^2 (up to constants), which is affine rather than proportional to α_ACS without the x_n^T x_n factor. The conclusion that the two criteria have the same greedy maximizer is still correct because both are monotone in x_n^T Σ_θ x_n, but the proportionality claim should be corrected.","section":"Section 4, Eq. (12)"},{"comment":"The sentence 'Empirically, we find that this property leads to smaller batches as more data points are acquired' is not supported by any experimental result in the paper; please provide evidence or remove the claim.","section":"Section 3, after Algorithm 1"},{"comment":"In Appendix B.2, the line after Eq. (20) writes Φ(ζ_m)Φ(ζ_m) where Φ(ζ_n)Φ(ζ_m) is intended; fix the typo.","section":"Appendix B.2"},{"comment":"The random-projection method in Section 5 uses J=10 throughout the experiments; since Eq. (17) is a Monte Carlo estimator, please report the sensitivity of the results to J or justify the choice.","section":"Section 5, Experiments"}],"recommendation":"major_revision","confidential_remarks":"The paper has a strong empirical component, but the central theoretical claim is currently unsupported. I would route the revision to an expert in coresets/active learning. The proportionality error in Section 4 should also be fixed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a genuinely useful application of Hilbert coresets to batch active learning. The paper does something new—it treats batch construction as sparse subset approximation of the expected complete-data log posterior and solves it with Frank-Wolfe. The closed forms for linear and probit regression are nice, and the connection to BALD and leverage scores is illuminating. The random-projection extension makes it work for any tractable likelihood, and the experiments are large-scale and mostly convincing. ACS-FW consistently beats random and is competitive with k-center/k-medoids, which is a reasonable result.\n\nNow the soft spots. The biggest one is exactly what the stress-test note flags: Section 3 asserts that a good function-space approximation implies the posterior will be close to the complete-data posterior, but no theorem or bound is given. This is not a minor technical detail. The quantity being approximated is the expectation over labels of the complete-data log posterior, while the actual posterior after labeling uses realized labels. Even an exact solution of the sparse-subset problem would not, by itself, guarantee near-optimality for the AL objective. So the 'principled' framing is overreach. The algorithm may still work as a diversity-promoting heuristic, and the experiments suggest it does, but the paper should either prove a transfer bound (like Campbell and Broderick do for coresets) or honestly reframe the contribution as heuristic.\n\nSecond, the analytic closed forms in Section 4 assume a Gaussian posterior with known noise variance. The experiments use an inverse Gamma prior on the noise variance, which yields a Student-t predictive posterior. The paper does not explain how the closed forms are applied in that setting. This is a real gap between theory and experiments.\n\nThird, the 'first principled, scalable and model-agnostic Bayesian batch AL approach' is a bit too strong; there is prior batch-AL work with a probabilistic flavor, and 'principled' is exactly what is in question.\n\nOn credit: source code is available, the experiments are fairly detailed, and the authors are honest about the weight-projection step and the variable batch size.\n\nBottom line: this is a solid, useful paper for the active learning and coresets communities. It deserves peer review, but the theory section needs either a proof or a more modest claim. If I were the editor, I would send it out and expect major revision.","headline":"A genuinely useful application of Hilbert coresets to batch active learning, with solid experiments and nice closed forms, but the 'principled' motivation rests on an unproven transfer assertion that the paper should either prove or soften.","tokens_in":16674,"tokens_out":2660,"would_cite":true,"duration_ms":29035,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62F15","68T05","62L05"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper frames batch active learning as sparse subset approximation of the expected complete-data log-posterior and shows that Frank-Wolfe optimization of this objective builds diverse, scalable query batches.","keywords":["batch active learning","Bayesian coresets","Frank-Wolfe optimization","sparse subset approximation","posterior approximation","random projections","acquisition functions","Bayesian linear regression"],"falsifier":"On a small regression problem where the full posterior is computable, run ACS-FW and a random batch of the same size, compute the posterior divergence between the batch-updated posterior and the complete-data posterior, and record $\\|L - L(w)\\|$ for both. If the random batch achieves as small a posterior divergence for a comparable function-space error, or if lower function-space error does not track lower posterior divergence, the motivating transfer claim is refuted.","tokens_in":15754,"feed_emoji":"🎯","tokens_out":7017,"duration_ms":68946,"temperature":0.7,"pith_summary":"This paper claims that batch active learning can be reframed as a sparse subset approximation: pick a small set of unlabeled points whose expected log-likelihood terms stand in for the entire pool's contribution to the Bayesian posterior. The authors solve the resulting optimization with Frank-Wolfe iterations under a Hilbert-space inner product, yielding an algorithm called ACS-FW. If the claim holds, Bayesian active learning no longer requires greedy single-point queries: ACS-FW produces diverse batches, scales linearly in the size of the unlabeled pool, and beats random selection on large regression and classification tasks. The price of admission is a heuristic step: the paper asserts, without a formal bound, that approximating the expected complete-data log-posterior in function space yields a posterior close to the full-data posterior.","feed_headline":"Frank-Wolfe turns batch active learning into sparse subset choice","feed_subtitle":"New ACS-FW algorithm picks diverse query batches that beat random selection at scale.","key_machinery":"The load-bearing object is the set of expected log-likelihood functions $L_m$ viewed as vectors in a function space with an inner product. The weighted Fisher inner product $\\langle L_n, L_m \\rangle_{\\hat{\\pi},F} = \\mathbb{E}_{\\hat{\\pi}}[\\nabla_\\theta L_n(\\theta)^T \\nabla_\\theta L_m(\\theta)]$ gives closed-form interpretable acquisition scores for linear and probit models; the weighted Euclidean inner product $\\mathbb{E}_{\\hat{\\pi}}[L_n(\\theta) L_m(\\theta)]$ only requires likelihood evaluations and can be estimated by random projections $\\hat{L}_n = J^{-1/2}[L_n(\\theta_1),\\dots,L_n(\\theta_J)]^T$ with $\\theta_j \\sim \\hat{\\pi}$. The Frank-Wolfe algorithm does the work of batch construction: at each iteration it selects the vertex of the polytope most aligned with the residual $L - L(w)$, then updates weights by line search, so the batch is built by diversity-promoting residual matching rather than by thresholding a fixed acquisition score.","core_discovery":"The central discovery is that the acquisition problem, which unlabeled points to label next, can be written as minimizing $\\|L - L(w)\\|$ in a Hilbert space, where $L_m(\\theta) = \\mathbb{E}_{y_m}[\\log p(y_m \\mid x_m, \\theta)] + H[y_m \\mid x_m, D_0]$ and $L = \\sum_m L_m$ is the expected full-pool log-likelihood under the current predictive posterior. Relaxing the binary selection vector $w$ to nonnegative weights on the polytope $\\{w : w_m \\ge 0,\\ \\sum_m w_m \\sigma_m = \\sigma\\}$ turns batch construction into a convex quadratic program that Frank-Wolfe solves by repeatedly adding the pool point whose log-likelihood vector is most aligned with the current residual. For linear regression the method yields the closed-form inner product $x_n^T x_m \\sigma_0^{-4} x_n^T \\Sigma_\\theta x_m$, connecting the greedy version to BALD and to leverage scores; for probit regression the closed form uses the bivariate Normal cdf and Owen's T function. For arbitrary models, $J$ random projections of the log-likelihood terms give an unbiased estimator of the Euclidean inner product, making the per-iteration cost linear in the pool size. Experiments on UCI regressions, CIFAR-10, SVHN, and Fashion MNIST show ACS-FW matching or exceeding existing batch methods and consistently beating random selection.","pith_inferences":["Editorial inference: the function-space objective is only a proxy; without a transfer bound between $\\|L - L(w)\\|$ and posterior divergence, the theoretical motivation could fail even if the algorithm works. Plugging in a coreset-style bound would be the natural repair.","Editorial inference: the final binarization of Frank-Wolfe weights discards the fractional weights that line search computed; keeping continuous weights or using them for weighted retraining might improve the selected batch.","Editorial inference: the inner product is a free design choice. Different kernels on $L_m$ would encode different notions of similarity between queries and could be tuned to a task, connecting the method to data geometry.","Editorial inference: the appendix's observation that variational inference beats MC Dropout suggests the gains of ACS-FW on classification may depend as much on the inference procedure as on the batch algorithm; a controlled study varying only the inference method would isolate this."],"forward_implications":["ACS-FW constructs batches that avoid the correlated queries of MAXENT and BALD, because the acquisition direction rotates after each selected point.","With random projections, batch construction runs in $O(|P| J)$ time, so the method scales to pools of hundreds of thousands of points.","On UCI regressions, CIFAR-10, SVHN, and Fashion MNIST, ACS-FW consistently beats random selection and performs at least as well as non-probabilistic batch AL baselines such as K-Center and K-Medoids.","For linear models, the greedy version of the acquisition score is proportional to BALD up to an input-magnitude factor, making the method a direct extension of existing information-theoretic AL.","The paper claims this provides a scalable, model-agnostic Bayesian batch active learning algorithm that does not require retraining after every single query."],"supporting_citations":[{"why":"Provides the Hilbert coreset sparse approximation and the Frank-Wolfe relaxation that ACS-FW adapts to active learning.","marker":"[10]"},{"why":"The Frank-Wolfe quadratic programming algorithm used to solve the relaxed batch selection problem.","marker":"[11]"},{"why":"Defines BALD, the greedy information-theoretic baseline ACS-FW is compared against and connected to.","marker":"[4]"},{"why":"Introduces the k-center coreset batch AL baseline and motivates the need to diversify query batches.","marker":"[5]"},{"why":"Earlier Bayesian coreset work that supplies the sparse log-likelihood approximation idea.","marker":"[9]"},{"why":"Neural linear model used in experiments to obtain a closed-form posterior over the last layer.","marker":"[19]"},{"why":"ResNet feature extractor used for the large-scale classification experiments.","marker":"[16]"},{"why":"Mean-field variational inference used to approximate the posterior in multiclass classification.","marker":"[37]"},{"why":"Supplies Owen's T function and the bivariate Normal identity used in the probit closed form.","marker":"[23]"}],"fun_headline_variants":["Frank-Wolfe batch selection goes sparse","Sparse subsets power batch active learning","Diverse batches via Bayesian sparse choice","Active learning at scale with Frank-Wolfe","Batch query selection via sparse approximation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a small approximation error in the expected complete-data log-posterior implies the corresponding batch posterior is close to the full-data posterior; the paper asserts this link but provides no theorem or bound for it.","fun_headline_variants_meta":{"raw":{"variants":["Frank-Wolfe batch selection goes sparse","Sparse subsets power batch active learning","Diverse batches via Bayesian sparse choice","Active learning at scale with Frank-Wolfe","Batch query selection via sparse approximation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000429,"raw_usage":{"total_tokens":2220,"prompt_tokens":996,"completion_tokens":1224,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":612,"completion_tokens_details":{"reasoning_tokens":1161}},"tokens_in":612,"tokens_out":1224,"duration_ms":11917,"temperature":1.0,"reasoning_tokens":1161,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:53:42.296176+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a small regression problem where the full posterior is computable, run ACS-FW and a random batch of the same size, compute the posterior divergence between the batch-updated posterior and the complete-data posterior, and record $\\|L - L(w)\\|$ for both. If the random batch achieves as small a posterior divergence for a comparable function-space error, or if lower function-space error does not track lower posterior divergence, the motivating transfer claim is refuted.","supporting_citations":[{"cited_title":"Automated scalable Bayesian inference via Hilbert coresets","cited_arxiv_id":null,"evidence_quote":"Provides the Hilbert coreset sparse approximation and the Frank-Wolfe relaxation that ACS-FW adapts to active learning."},{"cited_title":"An algorithm for quadratic programming","cited_arxiv_id":null,"evidence_quote":"The Frank-Wolfe quadratic programming algorithm used to solve the relaxed batch selection problem."},{"cited_title":"Coresets for scalable Bayesian logistic regression","cited_arxiv_id":null,"evidence_quote":"Earlier Bayesian coreset work that supplies the sparse log-likelihood approximation idea."},{"cited_title":"Deep Bayesian bandits showdown","cited_arxiv_id":null,"evidence_quote":"Neural linear model used in experiments to obtain a closed-form posterior over the last layer."},{"cited_title":"Deep residual learning for image recognition","cited_arxiv_id":null,"evidence_quote":"ResNet feature extractor used for the large-scale classification experiments."},{"cited_title":"Graphical models, exponential families, and variational inference","cited_arxiv_id":null,"evidence_quote":"Mean-field variational inference used to approximate the posterior in multiclass classification."},{"cited_title":"Tables for computing bivariate normal probabilities","cited_arxiv_id":null,"evidence_quote":"Supplies Owen's T function and the bivariate Normal identity used in the probit closed form."}],"review_version":1}