{"id":"c8851af0-999e-4567-aef5-9b18241d6f9b","arxiv_id":"2411.09250","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"SAAN reserves embedding space for future classes with assigned centers and adds a norm-distribution score to nearest-class-mean classifiers, improving few-shot class-incremental accuracy when plugged into existing methods.","lead":"This paper proposes a way to train few-shot class-incremental learning models by reserving chunks of feature space for future classes and by combining angle and norm information in the classifier. The method, called SAAN, is designed as a plug-in that raises final-session accuracy of existing incremental learning methods by about 2.6 to 3.3 percentage points on standard image benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The plug-in claim is not statistically established: main tables are single runs, hyperparameters are tuned on the same test benchmarks, and two of the six reported comparisons already fall below the advertised 3% improvement.","rationale":"The reader's weakest assumption is the log-normal shared distribution for embedding norms. That assumption is indeed weakly validated, but it is not the most load-bearing part of the central claim. The norm-distribution component contributes roughly one percentage point of accuracy (CIFAR100 ablation in Tab. 5: 52.72 versus 51.73; CUB200 in Tab. 6: 58.55 versus 57.56-57.72), and the compression coefficient C=0.005 makes the norm logits nearly flat, since p^0.005 is close to 1 for typical probabilities. Even if the distributional assumption failed completely, the plug-in gains would shrink rather than disappear. By contrast, the absence of repeated-run statistics directly threatens all six comparisons that constitute the plug-in claim, and the paper's own tables already contain two sub-3% results despite the contribution statement claiming 'over 3% across three datasets and two methods.' The paper is otherwise coherent, with a reasonable ablation structure and a clear separation of CCSA and ANJ contributions, so a conditional verdict remains appropriate; the condition should explicitly include multi-seed evaluation and a validation-based hyperparameter selection.","tokens_in":902,"tokens_out":985,"duration_ms":88882,"concrete_test":"Run the FACT+SAAN versus FACT and SAVC+SAAN versus SAVC comparisons on CIFAR100 and miniImageNet with at least 5 random seeds per condition, selecting hyperparameters on a held-out validation split and reporting mean plus/minus standard deviation and paired differences. If the mean improvement for FACT+SAAN on CIFAR100 remains at or below 0.5 percentage points, or if the 95% confidence interval for any of the six paired comparisons includes zero, the claim should be narrowed from 'over 3% across three datasets and two methods' to 'works in some settings with favorable tuning.'","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that SAAN, used as a plug-in, raises final-session accuracy by over 3% across three datasets and two methods (contribution 3). The supporting evidence in Tables 1-3 is six single-run paired comparisons, reported without error bars, seeds, or confidence intervals. Two comparisons already contradict the stated threshold: FACT+SAAN on CIFAR100 improves only 51.93 to 52.11 (+0.18 percentage points), and SAVC+SAAN on CIFAR100 improves 51.64 to 54.22 (+2.58 percentage points). Because FSCIL involves random session order and initialization, a single-run difference of +0.18 or +2.58 percentage points is within plausible seed noise. Moreover, the hyperparameters alpha=2, beta=0.4, and C=0.005 were selected by grid search on final-session accuracy of the same test benchmarks (Sec. 5.5, Figs. 7-8), so the reported gains may partly reflect test-set tuning rather than a robust plug-in property. Without repeated runs, a held-out validation split, or released code, the claim that SAAN 'consistently' improves both methods on all three datasets is not established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SAAN, a plug-in framework for few-shot class-incremental learning, consisting of two components: CCSA, which partitions the embedding space by assigning orthogonal class centers via the Hungarian algorithm and guiding embeddings with a cosine center loss, and ANJ, which models the log-norm distribution of each class and combines angle-based NCM logits with norm-based logits. The authors claim that SAAN achieves state-of-the-art performance and that adding it as a plug-in to FACT or SAVC improves final-session accuracy by over 3% across CIFAR100, CUB200, and miniImageNet. Experiments also cover an open-ended imbalanced CompCars setting, ablations, hyperparameter sensitivity, and a computational cost analysis.","tokens_in":142,"tokens_out":4071,"duration_ms":64099,"significance":"If the plug-in property were reliably established, this would be a practically useful contribution: the method is simple, adds negligible parameters and inference cost (Appendix D), and is designed to integrate with existing virtual-class FSCIL methods. The paper provides useful ablations showing the individual contributions of space allocation, the two-stage NCM, and norm distribution, and it includes some repeated experiments in the open-ended setting and for 2SNCM. However, the central quantitative claim is currently overstated and the main tables do not provide enough statistical evidence to support the claimed consistency of the plug-in gains.","major_comments":[{"comment":"The central claim that SAAN 'improves the final round accuracy by over 3% across three datasets and two methods' is not supported by the paper's own tables. In Table 3, FACT+SAAN improves from 51.93 to 52.11 on CIFAR100, which is +0.18 percentage points, and SAVC+SAAN improves from 51.64 to 54.22, which is +2.58 percentage points. Only the CUB200 and miniImageNet SAVC+SAAN comparisons exceed 3%. In addition, the abstract's 'state-of-the-art' statement is too strong: standalone SAAN is below NC-FSCIL and M2SD on all three datasets and below SAVC on CUB200 and miniImageNet. The claim should be corrected to report the actual per-dataset gains, or additional experiments should be provided if the 3% claim is meant to be an average or a best-case result.","section":"Abstract and Contribution 3, Tables 1-3"},{"comment":"The main plug-in comparisons are single runs reported without error bars, seeds, or confidence intervals. FSCIL results depend on random initialization and on the order of incremental sessions, so paired differences of +0.18 and +2.58 percentage points may be within run-to-run noise. The paper already uses repeated experiments in Section 5.3 and Appendix E, so the same practice should be applied to the main comparisons: report the mean and standard deviation over at least three to five seeds for FACT, SAVC, FACT+SAAN, and SAVC+SAAN on all three datasets. Without this, the claim that SAAN 'consistently' improves both methods on all datasets is not statistically established.","section":"Tables 1-3 and Section 5.1"},{"comment":"The hyperparameters alpha=2, beta=0.4, and C=0.005 are selected by grid search on the final-session accuracy of the same test benchmarks that are then used to produce the main results in Tables 1-3. This is a form of test-set tuning, and it weakens the claim that the plug-in gains reflect a general property of SAAN rather than per-dataset hyperparameter fitting. The authors should either use a held-out validation split for hyperparameter selection or show that a fixed, non-tuned hyperparameter setting yields qualitatively similar improvements across all datasets.","section":"Section 5.5, Figures 7 and 8"},{"comment":"The ANJ classifier relies on the assumption that log-norms are normally distributed for each base class and that all incremental classes share a single normal distribution. The evidence is limited to histograms and one Q-Q plot, and the authors themselves note that session 8 deviates from normality. Since this distributional assumption is a core component of ANJ, the paper should provide a quantitative goodness-of-fit test or a sensitivity analysis showing that the plug-in gains are not an artifact of the normality assumption. For example, the authors could compare the proposed parametric model against a nonparametric norm distribution or evaluate calibration of the norm logits.","section":"Section 4.2.2, Equation 10, Figure 10"}],"minor_comments":[{"comment":"M2SD is listed as a comparison method in Section 5.2 but is missing from Table 3 (CIFAR100); if results are unavailable, the paper should state this explicitly.","section":"Tables 1-3"},{"comment":"There is a typo: 'SSAN' appears in the conclusion where 'SAAN' is intended.","section":"Section 6"},{"comment":"The sentence 'which allows SSAN to achieve better performance' contains the same typo; it should read 'SAAN'.","section":"Section 5.2"},{"comment":"The table header repeats the same description for both 'Delta last' and 'Delta average'; the second should refer to average accuracy rather than 'last sessions'.","section":"Table 4"},{"comment":"The distribution name 'Patero' is a typo and should be 'Pareto', consistent with the text.","section":"Table 6"},{"comment":"In the related work text, 'MS2D' appears to be a typo for 'M2SD'.","section":"Section 2"},{"comment":"The harmonic mean is mentioned in the text but is not defined; please specify how it is computed and why it is used.","section":"Figure 6"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core plug-in idea is worth a look. SAAN is a straightforward combination: Hungarian-assigned orthogonal centers plus a cosine center loss to reserve space, and a norm-tail probability multiplied into cosine logits. The assembly is new, the ablations are clean, and the plug-in gains on CUB200 and miniImageNet are consistent and non-trivial (roughly +2 to +3 points final accuracy, with lower forgetting). The CompCars open-ended experiment uses three repeats and shows the same pattern. Credit where due: the write-up is honest about SAAN not beating SAVC on its own, and the ablation table gives a clear picture of what each piece contributes.\n\nThe soft spots are real but not fatal. The abstract's \"state-of-the-art\" claim is contradicted by the paper's own tables, where NC-FSCIL, M2SD, and SAVC are all higher on at least one benchmark. The \"over 3% improvement\" claim is simply not true for two of the six paired comparisons: FACT+SAAN on CIFAR100 is +0.18, SAVC+SAAN on CIFAR100 is +2.58. Those are single runs without error bars, so the plug-in property is not statistically established. Hyperparameters (alpha, beta, C, eta, lambda) were tuned on final-session accuracy of the same benchmarks, which is a real circularity concern. The log-normal assumption for embedding norms is load-bearing for the norm logits and is supported mainly by histograms and one Q-Q plot that shows a long tail in session 8. That assumption could break on other datasets.\n\nFor an active subfield, this is a useful plug-in paper rather than a breakthrough. It deserves peer review, conditional on code release, repeated runs, and a corrected abstract. I'd bring it to reading group because the space-allocation trick is simple enough to be worth discussing.","headline":"A clean plug-in method with credible gains on two benchmarks, but the advertised SOTA and 3% claims are overstated and the main tables lack error bars.","tokens_in":24355,"tokens_out":1588,"would_cite":false,"duration_ms":15379,"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":"SAAN claims a plug-in class-center space allocation and angle-norm classifier that raises final-session accuracy by over 3% in few-shot class-incremental learning.","keywords":["few-shot class-incremental learning","embedding space allocation","cosine center loss","angle-norm joint classifier","norm distribution","prototype learning","plug-in method","nearest class mean"],"falsifier":"Compute a Q-Q plot of feature-vector log lengths on the last incremental session of a new dataset, then compare SAVC with and without the norm-based score; if a visible deviation from the straight line coincides with the norm term adding no final-session gain, the normality assumption is the part that fails.","tokens_in":23288,"feed_emoji":"🧩","tokens_out":11472,"duration_ms":110611,"temperature":0.7,"pith_summary":"The paper tries to solve few-shot class-incremental learning (FSCIL), where a model must keep recognizing old classes while learning new ones from only a handful of examples. It proposes SAAN, a two-part scheme: CCSA partitions the embedding space by assigning each session a reserved subspace through orthogonal class centers and a cosine center loss, and ANJ replaces the cosine-only nearest-class-mean classifier with one that also models each class's embedding-norm distribution and combines angle and norm scores. The paper's central claim is that SAAN is competitive by itself and, more importantly, can be plugged into existing methods such as FACT and SAVC, raising final-session accuracy by over 3% on CIFAR100, CUB200, and miniImageNet while reducing forgetting. If that holds, existing feature-freezing FSCIL methods can be upgraded by adding SAAN's allocation loss and classifier, gaining accuracy on the final session without changing the frozen-feature paradigm.","feed_headline":"Space carving boosts few-shot incremental learning by over 3%","feed_subtitle":"Pre-assigned class centers and norm-aware scoring let existing models learn new classes without forgetting old ones.","key_machinery":"The machinery is the pair (CCSA, ANJ). CCSA is the space-allocating component: it fixes a set of orthogonal class centers, assigns them to classes via the Hungarian algorithm, and trains with a cosine center loss so that sample embeddings gather around their assigned center and away from others, with momentum center updates that decay so the reserved spaces survive. ANJ is the classifier component: it computes class means from normalized embeddings only in the base session (2SNCM), fits log-norm distributions to embedding lengths, and produces joint angle-norm logits by compressing the norm probability and multiplying it with the cosine logit. The argument depends on two mechanisms holding at once: new classes actually land in the reserved subspaces, and norm differences carry label information that the cosine score misses.","core_discovery":"The central discovery, on the paper's own terms, is that two weaknesses of FSCIL—old classes sprawling over the whole embedding space and nearest-class-mean classifiers ignoring vector length—can be addressed together. CCSA pre-generates mutually orthogonal class centers, matches them to classes with the Hungarian algorithm using cosine distance, and trains with a cosine center loss $L_{cc}=\\alpha L_1+\\beta L_2$ that pulls samples toward their own center and away from others, so each incremental session gets a reserved subspace. ANJ observes that embedding log norms differ between classes and sessions, models them as normal distributions (one per base class, one shared normal for all incremental classes), and forms joint logits $z_j = z^1_j (z^2_j)^C$, multiplying cosine similarity by a compressed norm probability. The experimental case is that SAAN alone beats FACT on all three benchmarks and that FACT+SAAN and SAVC+SAAN improve final-session accuracy by over 3% and lower forgetting.","pith_inferences":["Because ANJ only changes how a frozen embedding is scored, the norm-logit stage could be tested as a generic upgrade to other nearest-class-mean pipelines beyond FSCIL, provided the log-norm normality assumption is verified first.","The shared normal distribution for all incremental classes is a strong simplification; as an incremental run grows longer, fitting per-class or per-session norm distributions could recover information the shared prior throws away, and the paper's own session-8 Q-Q deviation points to where that limit would show up.","The Hungarian center assignment fixes the allocation once; a natural extension is to test whether allowing centers to be re-allocated when class semantics shift, or using hierarchical centers for very large class sets, changes the plug-in gain."],"forward_implications":["Plugging SAAN into FACT and SAVC lifts final-session accuracy by over 3% on CIFAR100, CUB200, and miniImageNet and reduces average forgetting by about 2.6 percentage points.","Used alone, SAAN beats FACT on all three standard benchmarks and beats SAVC on CIFAR100, showing the space allocation and norm classifier are competitive without contrastive pretraining.","The plug-in gains carry over to open-ended, imbalanced long-horizon settings: on CompCars with 20 incremental sessions, SAVC+SAAN exceeds SAVC by 2.83% in last-session accuracy.","Ablations show the largest single contribution comes from the cosine center pull that reserves space, with the norm-distribution term adding further gains, so both halves of the framework are claimed to be necessary for the full improvement."],"supporting_citations":[{"why":"SAAN embeds into and compares against this FACT virtual-prototype baseline.","marker":"[11]"},{"why":"SAAN embeds into this SAVC virtual-contrastive baseline and SAVC+SAAN outperforms it.","marker":"[16]"},{"why":"This NC-FSCIL fixed-prototype baseline motivates SAAN's semantic center assignment.","marker":"[12]"},{"why":"This TOPIC baseline establishes the FSCIL evaluation protocol used in the comparison tables.","marker":"[8]"},{"why":"This nearest-class-mean classifier supplies the cosine prediction rule that ANJ extends with norm logits.","marker":"[17]"},{"why":"This center loss is the basis for SAAN's cosine center loss for space allocation.","marker":"[51]"},{"why":"This Hungarian assignment algorithm matches class centers to classes in CCSA.","marker":"[52]"}],"fun_headline_variants":["Carve space for new classes, score by norm: +3% FSCIL","Subspace allocation plus norm-aware scoring lifts FSCIL","Angle-norm logits beat nearest-class-mean in FSCIL","Reserve subspace per class; multiply cosine by norm probability","3% boost: divide embedding space and model norm distributions"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the logarithms of feature-vector lengths follow a normal distribution—each base class with its own mean and variance, and all incremental classes sharing one mean and variance—so if the actual distribution is not normal on a new dataset, the norm-based scores become miscalibrated and the plug-in gains shrink.","fun_headline_variants_meta":{"raw":{"variants":["Carve space for new classes, score by norm: +3% FSCIL","Subspace allocation plus norm-aware scoring lifts FSCIL","Angle-norm logits beat nearest-class-mean in FSCIL","Reserve subspace per class; multiply cosine by norm probability","3% boost: divide embedding space and model norm distributions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001219,"raw_usage":{"total_tokens":5060,"prompt_tokens":1036,"completion_tokens":4024,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":652,"completion_tokens_details":{"reasoning_tokens":3934}},"tokens_in":652,"tokens_out":4024,"duration_ms":30043,"temperature":1.0,"reasoning_tokens":3934,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T20:50:59.625805+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute a Q-Q plot of feature-vector log lengths on the last incremental session of a new dataset, then compare SAVC with and without the norm-based score; if a visible deviation from the straight line coincides with the norm term adding no final-session gain, the normality assumption is the part that fails.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"SAAN embeds into and compares against this FACT virtual-prototype baseline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"SAAN embeds into this SAVC virtual-contrastive baseline and SAVC+SAAN outperforms it."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"This TOPIC baseline establishes the FSCIL evaluation protocol used in the comparison tables."},{"cited_title":"Mensink, J","cited_arxiv_id":null,"evidence_quote":"This nearest-class-mean classifier supplies the cosine prediction rule that ANJ extends with norm logits."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"This center loss is the basis for SAAN's cosine center loss for space allocation."}],"review_version":1}