{"id":"b30be381-cbdf-4de6-bbbf-1a0bca48477a","arxiv_id":"2411.17444","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"MSAL scores uncertainty by cosine similarity to fixed equiangular class prototypes, and MSAL-D adds prototype-based diversity, reporting improved AUBC on MNIST, SVHN, and TinyImageNet.","lead":"This paper proposes an active learning method that uses fixed equiangular points on a hypersphere as class prototypes for choosing which unlabeled images to annotate. It combines uncertainty and diversity without expensive clustering, and reports gains over several baselines on standard image benchmarks.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The equal-per-prototype diversity step is the load-bearing component of MSAL-D, yet it already hurts on CIFAR-10 and is never evaluated under class imbalance in Table 5.","rationale":"The paper's most concrete contribution is the two-stage MSAL-D selection; without it, MSAL is just another uncertainty metric. The evidence that the diversity stage works is mixed even in the paper's own tables: CIFAR-10 AUBC drops from MSAL to MSAL-D, while CIFAR-100 and TinyImageNet improve. Because no variance is reported, it is unclear whether these differences are meaningful. The imbalance experiments, the natural stress test for an equal-allocation rule, omit MSAL-D entirely. This makes the central claim conditional: the method may be sound, but the decisive experiment is missing. I agree with the reader's identification of the diversity heuristic but would sharpen it: the failure mode is not hypothetical; it is already visible on CIFAR-10, and Table 5 cannot rule it out for long-tailed data. I would therefore keep the CONDITIONAL verdict, since the core idea is plausible and the missing test is addressable.","tokens_in":122,"tokens_out":7135,"duration_ms":194380,"concrete_test":"Re-run Table 5 using MSAL-D (not only MSAL) on CIFAR-100-LT with imbalance factors 0.1 and 0.01, and on balanced CIFAR-10 with 5 seeds reporting mean ± std. Also instrument Algorithm 1 to count how often a prototype has fewer than b/C candidates in the pre-filtered pool and specify the fallback. If MSAL-D beats MSAL and Margin on the LT settings and does not lose to Margin on CIFAR-10, the diversity heuristic is supported; otherwise the central claim should be narrowed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that MSAL-D's prototype-based selection (Algorithm 1) provides diversity without clustering and outperforms other AL methods. The load-bearing step is the inner loop 'for c = 0 to C-1: find b/C closest samples,' which forces exactly b/C selections per fixed prototype. This rule is not tested where it is most fragile. In Table 2, on CIFAR-10 MSAL-D has AUBC 0.8107, below MSAL alone (0.8127) and below Margin (0.8147), so the diversity component can reduce performance. On TinyImageNet, the large gain (0.2717 vs 0.2443) may reflect crude class stratification rather than intra-class diversity; no diversity metric or per-prototype budget feasibility is reported. The long-tailed experiments in Table 5 report only MSAL, not MSAL-D, so the equal-allocation rule is never stress-tested under class imbalance, where many prototypes may have fewer than b/C candidates in the pre-filtered uncertain pool. Since the abstract's 'outperform across five benchmark datasets' rests on MSAL-D, this unverified heuristic is the weak point.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Maximally Separated Active Learning (MSAL), an uncertainty metric based on cosine similarity to fixed equiangular hyperspherical class prototypes, and MSAL-D, which adds diversity by pre-filtering the βb most uncertain samples and then selecting b/C samples closest to each prototype. The fixed prototypes come from the authors' prior maximum-separation work [33]. Experiments on MNIST, SVHN, CIFAR-10/100, and TinyImageNet compare against uncertainty-, diversity-, and combined-strategy baselines, with additional results on pseudo-labeling and long-tailed active learning. The central claims are that MSAL-D outperforms existing active learning methods across five benchmark datasets and that adding maximum separation boosts existing uncertainty strategies in both AUBC and final accuracy.","tokens_in":15715,"tokens_out":3699,"duration_ms":35215,"significance":"If the claims hold, the paper offers a simple and computationally cheap way to combine uncertainty and diversity without clustering, and it demonstrates that a fixed prototype head can stabilize representations throughout active learning. The work builds on a publicly available codebase and benchmarks against a standard suite of baselines, including BADGE, DBAL, and DeepAL+ implementations, which makes the empirical comparison reproducible in principle. The idea of using the same geometric structure both for uncertainty scoring and diversity allocation is elegant and potentially useful. However, the main significance is empirical, and the evidence as currently presented is mixed: several table entries contradict the stated conclusions, and the diversity component is not stress-tested under the class-imbalance setting where it is most fragile. The contribution is therefore promising but needs a more careful and honest empirical analysis before its claims can be accepted.","major_comments":[{"comment":"The claim that 'making the active learning approach maximally separated boosts the performance across the datasets in both AUBC and F-acc metrics' is directly contradicted by Table 3. For example, LeastConf(MSAL) on SVHN has AUBC 0.8623 versus 0.8653 for the non-maximally-separated LeastConf baseline, and on CIFAR-100 the AUBC is 0.4743 versus 0.4750. Several MSAL-D rows also decrease relative to their non-MSAL counterparts (e.g., Entropy(MSAL-D) on CIFAR-10 AUBC 0.8103 versus 0.8120). Additionally, the baseline values differ between Table 2 and Table 3: Margin on CIFAR-10 is reported as 0.8147 in Table 2 but 0.8133 in Table 3. The authors should correct the numbers, report a single consistent baseline set, and temper the claim to state where max separation helps and where it does not.","section":"Section 4.2, text after Table 3"},{"comment":"The equal-allocation diversity step ('for c = 0 to C-1: find b/C closest samples') is the load-bearing component of MSAL-D, yet it is never evaluated under class imbalance. Table 5 reports only MSAL, not MSAL-D, so we do not know how the rule behaves when many classes have fewer than b/C candidates in the pre-filtered βb uncertain set. This is a concrete failure mode: with imbalance factor 0.1 or 0.01, rare-class prototypes may have very few or no nearby samples in the uncertain pool, making the per-prototype budget infeasible. The paper should either specify a fallback for such cases or include MSAL-D results on the long-tailed benchmarks, and discuss whether the rule remains representative under skew.","section":"Section 3.2 and Algorithm 1"},{"comment":"The abstract claims outperforming 'various active learning methods across five benchmark datasets,' but on CIFAR-10 MSAL-D has AUBC 0.8107, which is below MSAL alone (0.8127) and below the Margin baseline (0.8147). Thus the diversity component actively hurts on one of the five datasets, and the overall claim is not supported. The authors should either restrict the claim to the datasets and metrics where MSAL-D is actually best, or analyze why the diversity step fails on CIFAR-10 (e.g., prototype-feature calibration, β choice, or per-class budget feasibility).","section":"Table 2, CIFAR-10 row and abstract"},{"comment":"All empirical results are reported as point estimates from three runs, with no standard deviations, confidence intervals, or significance tests. The differences that support the central claims are often small on the AUBC scale (e.g., 0.9820 vs 0.9787 on MNIST, 0.4780 vs 0.4777 on CIFAR-100), so without variance information a reader cannot judge whether the improvements are meaningful or noise. For a paper whose main contribution is empirical, the authors should report run-level variation (at minimum variance or error bars) and preferably a paired test across the common random initializations.","section":"Section 4.2 and Table 2/Table 3"}],"minor_comments":[{"comment":"There are several typos and grammatical errors, including 'from from' (Section 3.2), 'stragegy' (Section 2.1), 'evalution' (Section 4), 'unlabled' (Section 2.1), 'ubiqutous' (Section 2.2), and 'the the' (Section 5). A careful proofreading pass is needed.","section":"Throughout"},{"comment":"The caption reads 'pre-filter factor on MNIST and SVHN datasets,' but the right panel is labeled 'pre-filter factor on CIFAR10.' The caption should match the figure contents.","section":"Figure 4 caption"},{"comment":"The inner loop uses 'arg max_c α_MSAL' but the surrounding text and Figure 2 refer to selecting samples 'closest to each cluster center.' Since α_MSAL is a similarity score, 'arg max' and 'closest' are consistent, but this should be stated explicitly to avoid confusion; also 'for each class C' mixes the loop variable and the total number of classes.","section":"Algorithm 1"},{"comment":"There is an extra space in some entries, e.g., '0 .9889' and '0 .9142,' which should be removed.","section":"Table 4"},{"comment":"The radius ρ is chosen as 0.1 for MNIST and 1 for other datasets 'using the heuristics as defined in [33],' but the paper does not report any sensitivity analysis for ρ. Since ρ scales the logits and directly affects the uncertainty scores, a brief ablation or justification would strengthen the claim that the method is not overly sensitive to this inherited hyperparameter.","section":"Section 4.4"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a reasonable empirical study of a simple idea, but the presentation currently overstates the results: the text after Table 3 and the abstract do not match the reported numbers, and the diversity component is not tested in the imbalance setting where its equal-allocation rule is most likely to break. The dependence on the authors' own prior work [33] is disclosed and the construction is clear, so this is not a novelty concern, but the empirical claims need to be made precise and the missing experiments or analyses added before the paper can be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know before you skim this one. First, it is a clean, simple transfer of a known construction: fixed equiangular hyperspherical prototypes [33] are used as the final classifier layer during active learning training, uncertainty is measured as cosine similarity to the nearest prototype (MSAL), and diversity comes from pre-filtering uncertain samples and then taking equal numbers closest to each prototype (MSAL-D). Second, the empirical story is overclaimed. The abstract says 'outperform over various active learning techniques across five benchmark datasets,' but Table 2 shows MSAL-D only wins on MNIST, SVHN, and TinyImageNet. On CIFAR-10 it is below Margin and below MSAL alone; on CIFAR-100 it is not the best.\n\nCredit where due: the combination is new, the method is genuinely easy to drop into an existing pipeline, and the gains on three datasets are real. The fixed prototype head also improves several standard uncertainty metrics in many settings (Table 3), which is useful for people who want a cheap diversity mechanism without clustering. Code is available, and the paper is written without obfuscation.\n\nThe soft spots are real and fixable. The text overclaims twice: after Table 3 it says maximum separation boosts performance across all datasets in both metrics, but LeastConf(MSAL) on SVHN and CIFAR-100 is slightly worse in AUBC. Before Table 4 it says MSAL outperforms CEAL across all settings, but CIFAR-100 F-acc is lower. There are no variance bars despite averaging three runs. The pre-filter factor beta is tuned on the evaluation curves of the same benchmarks. MHEAL, a closely related hyperspherical AL method, is omitted. The runtime statement ('6-10% more than entropy') is made without a protocol.\n\nThe load-bearing weakness is the diversity step. Forcing exactly b/C samples per prototype is a strong assumption. It works on TinyImageNet but already underperforms on CIFAR-10, and it is never tested under class imbalance: Table 5 only reports MSAL, not MSAL-D. No diversity metric or per-prototype budget feasibility is reported. That's the place to focus a revision.\n\nBottom line: the core idea survives, but the paper needs a serious round of corrections. A good referee can help. I'd take it, not desk-reject it.","headline":"Simple repurposing of fixed hyperspherical prototypes for active learning; the core idea is real, but the 'wins across five datasets' claim overstates the tables and the diversity step is untested where it should hurt most.","tokens_in":16431,"tokens_out":3409,"would_cite":true,"duration_ms":31383,"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":"Using fixed equiangular class prototypes as the network head, this paper shows that active learning gets both uncertainty and diversity without clustering, outperforming several baselines across five image benchmarks.","keywords":["active learning","hyperspherical prototypes","uncertainty sampling","diversity sampling","maximum separation","image classification","long-tailed classification","pseudo-labeling"],"falsifier":"The paper's own long-tailed results provide a check: on CIFAR-100 with imbalance factor 0.1, MSAL's AUBC of 0.3230 falls below margin sampling's 0.3263. A controlled run that varies the per-prototype allocation from equal ($b/C$) to class-frequency-proportional, keeping all other settings fixed, would settle whether the equal-allocation diversity heuristic is what hurts under imbalance and whether a simple correction restores the claimed edge.","tokens_in":15250,"feed_emoji":"🎯","tokens_out":8570,"duration_ms":71562,"temperature":0.7,"pith_summary":"Active learning chooses which unlabeled examples to label next, typically by picking the model's most uncertain samples, but that tends to select similar, redundant examples. This paper argues that fixing a set of maximally separated (equiangular) points on the hypersphere as class prototypes solves both sides of the problem at once: the prototypes act as an inductive bias that stabilizes class representations, and cosine similarity to the nearest prototype doubles as an uncertainty score. To add diversity without clustering, the method pre-filters the most uncertain examples and then takes an equal number of those closest to each prototype, treating prototypes as cluster centers. The paper reports that this strategy, MSAL-D, achieves the best overall accuracy-budget scores on MNIST, SVHN, and TinyImageNet, and that adding the fixed prototype head improves several classic uncertainty metrics on most settings. A sympathetic reader would take the central claim to be that hyperspherical uniformity can replace separate clustering and representation-learning machinery in active learning.","feed_headline":"Fixed prototype points replace clustering in active learning","feed_subtitle":"Using fixed equiangular class vectors, the method beats clustering-based samplers on three of five benchmarks.","key_machinery":"The carrying object is the fixed equiangular prototype matrix $P_c \\in \\mathbb{R}^{(C-1) \\times C}$ built by the maximum-separation construction (the recursive formula with vectors separated by angle $-1/k$), used as the network's fixed last layer so that logits are $\\rho P_c^T \\Phi(x)$. The uncertainty score $\\alpha_i^{\\mathrm{MSAL}} = \\max_c ( \\hat{x}_i \\cdot P_c^T ) / \\max \\| \\hat{x}_i \\|_2$ measures how far a sample sits from every prototype; the diversity step treats the $C$ prototypes as permanent cluster centers and allocates the batch by taking $b/C$ closest samples per prototype from the pre-filtered $\\beta b$ most uncertain samples. The paper's key mechanism is that a single fixed geometry does both jobs: it shapes the learned embedding space through training and it partitions the uncertain pool for batch selection.","core_discovery":"Writing as the authors would: enforcing maximum separation through fixed class prototypes in the network's last layer yields stable, well-separated embeddings from the first active learning round onward, and those same prototypes give a simple, cluster-free way to select both uncertain and diverse samples. Uncertainty is scored as the cosine similarity between a sample's feature and its closest prototype; samples far from all prototypes are most uncertain. Diversity is obtained by retaining the $\\beta b$ most uncertain samples, then selecting the $b/C$ samples closest to each of the $C$ prototypes, so the queried batch spreads across the class-separated regions of the space. On five image benchmarks (MNIST, SVHN, CIFAR-10, CIFAR-100, TinyImageNet) the combined strategy MSAL-D reports the highest area-under-the-budget-curve (AUBC) on three of them, and swapping in other uncertainty metrics (least confidence, margin, entropy) under the fixed-prototype head instead of a trainable head generally improves their results.","pith_inferences":["Beyond the paper: the diversity step's equal per-prototype allocation effectively enforces class balance in the query batch, so on highly imbalanced pools the number of rare-class samples that survive the uncertainty pre-filter will determine whether the method oversamples or starves those classes; this could be tested by varying $\\beta$ per class.","Beyond the paper: the t-SNE stability across rounds suggests the fixed-prototype head may also preserve geometry in continual or semi-supervised learning, where the embedding space must not drift between data arrivals; an experiment would be to swap the uncertainty score for a small supervised signal and watch whether clusters stay fixed.","Beyond the paper: since the prototypes only need a fixed set of landmarks, the same uncertainty-plus-diversity recipe could be applied to pixel-level queries in segmentation or to object proposals in detection, which the paper lists as future work but does not test."],"forward_implications":["MSAL-D reports the highest AUBC on MNIST (0.9820), SVHN (0.8657), and TinyImageNet (0.2717), beating the clustering-based combined strategies BADGE and DBAL on three of the five datasets.","Adding the fixed prototype head to least confidence, margin, and entropy sampling improves or matches their AUBC on most settings in Table 3, so the diversity and uncertainty gains are not tied to the specific MSAL score.","Because the diversity step uses fixed prototypes instead of clustering, runtime stays only 6–10% above entropy sampling.","Pseudo-labeling with the prototype-based confidence (CEAL-MSAL) surpasses the entropy-based CEAL on all five datasets in AUBC and in final accuracy.","On long-tailed CIFAR-10 with imbalance factor 0.01, MSAL achieves the best final accuracy among compared uncertainty methods, though on long-tailed CIFAR-100 with imbalance factor 0.1 it trails margin sampling."],"supporting_citations":[{"why":"Supplies the fixed equiangular prototype matrix used as the network head and as the cluster centers in the diversity step.","marker":"[33]"},{"why":"Contributes the pre-filter of the most uncertain examples that MSAL-D adapts before its prototype-based allocation.","marker":"[92]"},{"why":"Provides the evaluation toolkit and the baseline implementations whose numbers are marked with dagger in the paper's tables.","marker":"[88]"},{"why":"Defines the AUBC metric and the benchmark survey from which several baseline results are taken.","marker":"[87]"},{"why":"BADGE is the main clustering-based combined baseline the paper claims to outperform on three datasets.","marker":"[2]"},{"why":"CEAL is the pseudo-labeling baseline that MSAL with prototype-based confidence exceeds in Table 4.","marker":"[74]"},{"why":"Supplies the class-imbalance settings used in the long-tailed active learning experiments in Table 5.","marker":"[14]"}],"fun_headline_variants":["Fixed class prototypes banish clustering from active learning","Active learning with fixed hyperspherical prototypes beats clustering","No-cluster active learning: fixed prototypes for uncertainty and diversity","MSAL: Equiangular prototypes make active learning maximally separated","Three of five benchmarks: fixed-prototype active learning tops clustering"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The diversity step assumes that the learned features are calibrated around the fixed prototypes, so partitioning the pre-filtered uncertain pool by nearest prototype and taking $b/C$ samples per class yields a representative and diverse batch.","fun_headline_variants_meta":{"raw":{"variants":["Fixed class prototypes banish clustering from active learning","Active learning with fixed hyperspherical prototypes beats clustering","No-cluster active learning: fixed prototypes for uncertainty and diversity","MSAL: Equiangular prototypes make active learning maximally separated","Three of five benchmarks: fixed-prototype active learning tops clustering"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000172,"raw_usage":{"total_tokens":1230,"prompt_tokens":858,"completion_tokens":372,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":474,"completion_tokens_details":{"reasoning_tokens":290}},"tokens_in":474,"tokens_out":372,"duration_ms":4146,"temperature":1.0,"reasoning_tokens":290,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:06:32.586839+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"The paper's own long-tailed results provide a check: on CIFAR-100 with imbalance factor 0.1, MSAL's AUBC of 0.3230 falls below margin sampling's 0.3263. A controlled run that varies the per-prototype allocation from equal ($b/C$) to class-frequency-proportional, keeping all other settings fixed, would settle whether the equal-allocation diversity heuristic is what hurts under imbalance and whether a simple correction restores the claimed edge.","supporting_citations":[{"cited_title":"In: Advances in Neural Information Processing Systems (2022)","cited_arxiv_id":null,"evidence_quote":"Supplies the fixed equiangular prototype matrix used as the network head and as the cluster centers in the diversity step."},{"cited_title":"In: IJCAI","cited_arxiv_id":null,"evidence_quote":"Defines the AUBC metric and the benchmark survey from which several baseline results are taken."},{"cited_title":"IEEE Transactions on Circuits and Systems for Video Technology27(12), 2591–2600 (2016)","cited_arxiv_id":null,"evidence_quote":"CEAL is the pseudo-labeling baseline that MSAL with prototype-based confidence exceeds in Table 4."}],"review_version":1}