{"id":"29419d33-5f12-4730-8dda-2384bb1d139a","arxiv_id":"2412.13714","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"AnchorInv uses feature-space anchor points to invert synthetic replay samples, improving few-shot class-incremental learning on physiological time series without storing raw data.","lead":"This paper proposes AnchorInv, a method that generates synthetic training samples from stored feature-space \"anchor points\" to help few-shot class-incremental learning avoid forgetting on physiological signals. It reports state-of-the-art results on three EEG and EMG datasets, with a privacy-motivated alternative to storing raw data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Inverted replay samples are visually off-manifold (Fig. 5), yet the RealReplay upper bound exists only for BCI (Table 8); without it for NHIE/GRABMyo, the faithful-replay mechanism behind AnchorInv's forgetting prevention is unverified.","rationale":"The reader's weakest_assumption aligns with mine: the inversion may not produce on-manifold samples. The reader notes Figure 5 and the missing RealReplay on NHIE/GRABMyo. I agree. However, I also note a secondary concern from Appendix D: test-set-based early stopping is disclosed for baselines but the same procedure is not explicitly stated for AnchorInv; this should be clarified. Nevertheless, that issue is less load-bearing than the off-manifold question because even a perfect evaluation protocol would not establish the mechanism if the replay samples are not representative. The RealReplay comparison on one dataset is encouraging, but it is insufficient to generalize to all three. My recommendation is to keep the reader's CONDITIONAL verdict: the empirical results may hold, but the central mechanism needs either further evidence (RealReplay on other datasets) or a revised claim. Since my concern doesn't move the verdict, I set verdict_should_be to UNCHANGED.","tokens_in":18532,"tokens_out":12837,"duration_ms":118896,"concrete_test":"Run the RealReplay baseline from Table 8 on the NHIE and GRABMyo settings: store the same number of real base-class samples as anchors used (e.g., 50 per class) plus all incremental support samples, finetune with Lnew + Lold exactly as AnchorInv, and compare the final-session Macro-F1 to AnchorInv. If AnchorInv is substantially worse than RealReplay on either dataset (beyond the ~0.1 point gap seen on BCI), the inversion is not producing replay samples that carry the information needed to preserve old classes, calling the mechanism into question.","verdict_should_be":"UNCHANGED","load_bearing_attack":"AnchorInv's core premise (Eq. 11 + Eq. 13) is that minimizing MAE between the embedding of a synthetic input and a stored anchor produces replay samples representative of old classes, so that finetuning on Lold preserves prior knowledge. The paper's own Figure 5 contradicts the representativeness of the samples: real and inverted EEG are visually distinct, with inverted samples dominated by high-frequency fluctuations, even though feature-space MAE is only 0.06-0.08. This suggests the inversion finds off-manifold 'adversarial-like' inputs that match the anchor embedding without lying on the true data manifold. The only direct check that such samples are as useful as real data is the RealReplay upper bound in Table 8, and it is reported only for BCI. NHIE and GRABMyo—where the paper claims SOTA—have no such verification. If the inverted samples are not faithful, the improvement over baselines could come from a different mechanism (e.g., noise-based regularization), and the stated contribution of 'faithful representation' and 'knowledge forgetting prevention' is not supported by the presented evidence. This is load-bearing because the central claim is not just that AnchorInv scores well, but that it does so by generating representative replay samples from feature-space anchors.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes AnchorInv, a buffer-replay method for few-shot class-incremental learning (FSCIL) on physiological time series. Instead of storing raw samples, AnchorInv stores selected feature-space anchor points from each session and uses model inversion—minimizing mean absolute error between the embedding of a synthetic input and a stored anchor (Eq. 11)—to generate replay samples. These synthetic samples are then used together with the new few-shot samples to finetune the backbone (Eqs. 12–14). The authors evaluate on three public datasets (BCI-IV 2a, NHIE, GRABMyo) with 1-way-10-shot incremental sessions, reporting Macro-F1 across multiple random trials. They report that AnchorInv achieves the best Macro-F1 on all datasets across all sessions, with statistically significant differences from baselines, and present ablations on anchor selection, number of anchors, number of shots, and base-class count. The paper also claims the approach protects data privacy because raw data are not stored.","tokens_in":18836,"tokens_out":2526,"duration_ms":23274,"significance":"The paper targets a real and under-studied setting: FSCIL for physiological signals with very few base classes, where the authors argue that backbone finetuning is necessary. The proposed idea—using feature-space anchors to guide inversion for replay—is reasonable and distinct from label-space inversion baselines such as DeepDream and DeepInv. The manuscript has several strengths: it performs multi-trial evaluation with reported means and standard deviations, applies Wilcoxon signed-rank tests, compares against a range of baselines, and includes ablations on anchor selection, anchor count, shot count, and base-class count. If the central mechanism is verified, the approach would be a practical contribution to privacy-conscious incremental learning for biomedical data. However, the paper's core evidence that the inverted samples are representative of old classes is incomplete, and there are unresolved questions about the fairness of the hyperparameter selection procedure relative to baselines.","major_comments":[{"comment":"The RealReplay upper bound—the direct check that synthetic replay samples carry information comparable to real samples—is reported only for BCI (Table 8). For NHIE and GRABMyo, where the paper claims state-of-the-art results (Tables 2b and 2c), no such verification is provided. Since the central claim is that AnchorInv preserves prior knowledge by generating representative replay samples, the absence of this check leaves open the alternative explanation that the improvements arise from a generic regularizer rather than from faithful replay. Please provide RealReplay comparisons for NHIE and GRABMyo, or otherwise experimentally demonstrate that finetuning on the inverted samples transfers to held-out real data of the base classes.","section":"§4.3, Appendix E.2 (Table 8)"},{"comment":"The baseline finetuning schedules are explicitly selected on the test set: Appendix D states that the number of iterations is chosen by running 10 trials and selecting the best-performing iteration based on test-set performance. In contrast, Section C.2 reports AnchorInv's finetuning iteration ranges (e.g., 1000–1550 for BCI) without stating whether these were selected on the test set or by a validation criterion. If AnchorInv's schedule was tuned on test performance while baseline schedules were chosen in the same way, this should be stated; if AnchorInv's schedule was chosen differently, the comparison may be unfair. Please clarify the selection procedure and, if needed, report results under a unified validation protocol.","section":"§C.2 and Appendix D"},{"comment":"The qualitative evidence in Figure 3 that inverted feature vectors overlap with anchors is guaranteed by the MAE objective in Eq. (11) and therefore cannot independently support the representativeness of the inverted samples. The paper itself notes in Section 4.3 (Qualitative Analysis) and Figure 5 that real and inverted samples are visually distinct and that inverted samples are dominated by high-frequency fluctuations. This suggests the inversion may find off-manifold inputs that match the anchor embedding without lying on the true data manifold. Please add a quantitative test of representativeness, for example: (a) train a classifier on inverted samples and evaluate on real base-class test data, or (b) measure whether finetuning with Lold on inverted samples changes the model's embeddings of real base-class inputs in the intended way. Such a test would directly support the claim that the replay samples are 'faithful and representative' rather than merely matching the anchors in feature space.","section":"§3.2, Eq. (11), Figures 3 and 5"},{"comment":"The claim that AnchorInv 'protects privacy' and 'adheres to data privacy and sharing regulations' is asserted without a formal privacy analysis or a well-defined threat model. Storing feature-space anchors and generating synthetic samples from them reduces the direct exposure of raw data, but the paper demonstrates that anchors can be inverted to produce samples, so the privacy guarantee is not automatic. Please either provide a formal statement (e.g., under what adversary model the stored anchors cannot be used to recover identifiable raw signals) or soften the claim to a practical mitigation rather than a guarantee. This is load-bearing because privacy preservation is stated as a core motivation and contribution.","section":"Abstract, §1, §3.2"}],"minor_comments":[{"comment":"The reproducibility checklist states that all source code implementing the new method is included and will be made publicly available, while Appendix H says only that code will be made available upon publication. Please align these statements so the reader knows what is available at submission time.","section":"Appendix H and Reproducibility Checklist"},{"comment":"There are typographical errors: 'GRAMyo' in the Table 7 caption should be 'GRABMyo', and 'Distrbution' in the captions of Figures 7 and 8 should be 'Distribution'.","section":"Table 7 and Appendix G"},{"comment":"The sentence 'Across all incremental sessions, AnchorInv is statistically significantly different from all baseline methods' is ambiguous because it does not state the pairing used in the Wilcoxon test (e.g., paired per trial and session) or whether a multiple-comparison correction was applied. Please clarify the test protocol.","section":"§4.3 and Table 2"},{"comment":"The base-session performance differs among methods because Finetune and NC-FSCIL use an MLP classifier while the other methods use a metric classifier; the paper explains this, but the base-session rows in Table 2 are not directly comparable. It would be helpful to state this explicitly in the table caption or the main text.","section":"§C.1 and §C.2"}],"recommendation":"major_revision","confidential_remarks":"The paper's core idea is promising and the empirical setup is more careful than most FSCIL papers. The main issue is not the method's plausibility but the lack of direct evidence that the inverted samples are as faithful as real replay samples, especially on the two datasets where no RealReplay bound is provided; the privacy claim also needs to be tempered or formalized. These are fixable with additional experiments and clarifications, so I recommend major revision rather than rejection. I would also suggest the editor check whether the reproducibility checklist overstates code availability relative to Appendix H."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear [Colleague],\n\nPunchline: AnchorInv is a genuinely useful empirical paper for FSCIL on physiological time series, but the \"faithful representation\" claim is the weakest part and needs either more evidence or softer wording.\n\nWhat's actually new: The idea of guiding model inversion with feature-space anchors rather than label-space objectives, applied to replay in FSCIL, is a sensible variation on DeepInv/DeepDream. The multi-trial evaluation protocol (M random few-shot sets per session) is a good methodological contribution—it addresses real variability in few-shot FSCIL benchmarks. The paper also does a nice job showing that with very few base classes, frozen-backbone methods struggle, which motivates finetuning. The experiments are reasonably thorough: three datasets, multiple ablations (anchor selection, number of anchors, number of shots, base class subsets), and statistical tests with multiple trials. The code isn't out yet, but the pseudocode and hyperparameters are detailed enough to reproduce.\n\nSoft spots, in order of severity:\n\n1. The finetune schedule ambiguity. Appendix D says baselines pick iterations by test-set performance across 10 trials. For AnchorInv, the main text just gives ranges (e.g., 1000–1550 iterations) without saying whether the same test-set selection was used. If your method's schedule was also tuned on test, say so; if not, the comparison is uneven. This is fixable but needs clarification.\n\n2. The \"faithful representation\" evidence is thin. Figure 5 shows inverted EEG samples are visually dominated by high-frequency fluctuations and look quite different from real signals, even though feature-space MAE is 0.06–0.08. That's okay if your claim is just \"the inversion hits the anchor,\" but the paper calls them \"representative\" and \"faithful.\" The BCI RealReplay results (Table 8) show AnchorInv matches real-sample replay, which supports the practical usefulness. But that check is missing for NHIE and GRABMyo, where the SOTA claims are made. Without those upper bounds, you can't rule out that the gains come from some regularizing side effect rather than from preserving the actual class distribution. Add RealReplay (or at least a real-sample replay ablation) for the other datasets before making the representativeness claim.\n\n3. Privacy claim is asserted, not analyzed. Inversion from feature anchors could in principle leak information about training data. A short discussion of what privacy model you're assuming would suffice.\n\nThese are all addressable, and none of them kills the empirical result. The multi-trial evaluation and the anchor-inversion idea are worth having in the literature. The paper deserves a serious referee; I'd send it to review, and I'd ask for the clarifications above rather than rejecting.\n\nFor who: If you work on incremental learning for biosignals or on replay-without-data methods, this is a useful read. I'd cite it as a strong empirical baseline.\n\nRecommendation: engage with it, but insist on the RealReplay completeness and the finetuning description.","headline":"Solid empirical FSCIL paper with a useful anchor-inversion trick and a good multi-trial protocol, but the 'faithful replay' claim needs more evidence and the finetune tuning is ambiguous.","tokens_in":19326,"tokens_out":3696,"would_cite":true,"duration_ms":32195,"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":"AnchorInv stores feature-space anchor points instead of raw signals, inverts them into replay samples, and reports state-of-the-art few-shot incremental learning results on three EEG and EMG datasets.","keywords":["few-shot class-incremental learning","model inversion","physiological time series","EEG","EMG","catastrophic forgetting","buffer replay","feature space"],"falsifier":"Measure the last-session Macro-F1 on NHIE and GRABMyo when the finetuning regularization uses real stored samples instead of inverted ones, a real-replay upper bound; if AnchorInv falls clearly below that bound on either dataset, the claim that inverted samples carry as much prior-class information as real samples is contradicted. A second check: replace the anchor-guided mean-absolute-error inversion loss with a label-space cross-entropy inversion loss; if performance does not drop, feature-space guidance specifically is not the cause of the reported gains.","tokens_in":18358,"feed_emoji":"🧠","tokens_out":12508,"duration_ms":94187,"temperature":0.7,"pith_summary":"This paper sets out to establish that a replay-based few-shot class-incremental learning method can work for physiological time series even when the base training set has very few classes. Instead of storing raw signals, AnchorInv stores a small set of anchor points in the feature space, synthesizes replay samples by inverting those anchors, and finetunes the backbone on the synthetic samples together with the new few-shot examples. The paper argues that this preserves old-class knowledge while adapting to new classes and respects privacy, and it reports that the method outperforms state-of-the-art baselines across three EEG and EMG datasets. The paper also argues that existing FSCIL evaluations are too sensitive to which few-shot samples happen to be drawn, so it repeats adaptation over many randomly sampled support sets to report mean and standard deviation.","feed_headline":"Feature-anchor replay beats baselines in few-shot EEG/EMG learning","feed_subtitle":"It beats prior methods on all incremental sessions across three EEG and EMG benchmarks, while storing no raw signal data.","key_machinery":"The carrying mechanism is the anchor set: a small collection of embedding vectors, one or more per base class, selected by random sampling or clustering from the feature distribution of the training set. At each incremental session these anchors are the inversion targets; minimizing mean absolute error between a randomly initialized input's embedding and an anchor yields a synthetic replay sample, and the replay set is used in a finetuning loss that regularizes old-class knowledge while the backbone adapts to new classes. The metric classifier is a prototype-based softmax over negative cosine distances, so keeping old-class embeddings near their anchors is what prevents the feature space from shifting when new classes are integrated.","core_discovery":"The central claim is that guiding model inversion by feature-space anchor points, rather than by label-space classification loss, produces replay samples that prevent catastrophic forgetting more effectively during few-shot incremental sessions. Concretely, after base training the method computes a set of anchor points summarizing each base class's embeddings, solves an optimization to synthesize samples whose embeddings match those anchors under mean absolute error, and then finetunes using a loss that combines cross-entropy on the new few-shot samples with a regularization term on the inverted old-class samples. The paper reports that this achieves the best Macro-F1 across all incremental sessions on BCI, NHIE, and GRABMyo, with statistically significant differences from baselines at the p < 0.05 level by a two-sided signed-rank test, and that its performance closely tracks a real-replay upper bound on BCI.","pith_inferences":["Because inversion only needs the embedding network and stored anchors, the same procedure should transfer to other physiological modalities or restricted-data settings; testing on ECG or wearable motion signals would be a natural next step.","The visual distinctness of inverted samples from real signals suggests AnchorInv may be regularizing the feature space rather than reproducing faithful waveforms, meaning its success does not depend on replay samples looking realistic.","The paper only tests the real-replay upper bound on BCI; replicating that comparison on NHIE and GRABMyo would show whether the near-RealReplay result generalizes or is specific to within-subject BCI data."],"forward_implications":["Incremental learning on physiological signals can be privacy-preserving: no raw base-session samples need to be stored, only anchor points in feature space.","The method is most valuable exactly when base classes are scarce, because it finetunes the backbone instead of relying on a frozen generalizable feature extractor.","Reporting mean and standard deviation across multiple randomly sampled few-shot support sets gives a more stable expectation of real-world performance than a single split.","On BCI, AnchorInv's final performance is close to that of replaying real stored samples, suggesting the synthetic samples carry comparable regularization information."],"supporting_citations":[{"why":"DeepDream: inversion-based generation approach that AnchorInv extends and uses as a comparison baseline.","marker":"Mordvintsev, Olah, and Tyka 2015"},{"why":"DeepInv: establishes model inversion for data-free knowledge transfer; AnchorInv adapts this idea to feature-space targets.","marker":"Yin et al. 2020"},{"why":"EEG-Conformer: provides the convolutional-transformer backbone used for all experiments.","marker":"Song et al. 2023a"},{"why":"ProtoNet: supplies the metric/prototype classifier formulation and a baseline that relies on a strong feature extractor.","marker":"Snell, Swersky, and Zemel 2017"},{"why":"NC-FSCIL: a state-of-the-art vision FSCIL baseline re-implemented for physiological data.","marker":"Yang et al. 2023"},{"why":"TEEN: a second state-of-the-art vision FSCIL baseline used in the comparison.","marker":"Wang et al. 2023"},{"why":"BCI Competition IV 2a dataset: one of the three evaluation benchmarks.","marker":"Brunner et al. 2008"},{"why":"NHIE dataset: the neonatal EEG grading benchmark.","marker":"O'toole et al. 2023"},{"why":"GRABMyo dataset: the EMG gesture benchmark.","marker":"Jiang, Pradhan, and He 2022"}],"fun_headline_variants":["Anchor-guided replay tops few-shot physiological learning","No raw data needed: anchor replay wins on EEG/EMG","Feature anchors beat baselines in incremental biosignal learning","Synthetic replay via anchors curbs forgetting on few-shot EEG","Anchor inversion beats raw replay in few-shot biosignal learning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that making a synthetic signal map to the same internal representation as an old class also makes that synthetic signal useful for preserving the old class during later training, even if the synthetic signal does not look like a real one.","fun_headline_variants_meta":{"raw":{"variants":["Anchor-guided replay tops few-shot physiological learning","No raw data needed: anchor replay wins on EEG/EMG","Feature anchors beat baselines in incremental biosignal learning","Synthetic replay via anchors curbs forgetting on few-shot EEG","Anchor inversion beats raw replay in few-shot biosignal learning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000772,"raw_usage":{"total_tokens":3405,"prompt_tokens":919,"completion_tokens":2486,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":535,"completion_tokens_details":{"reasoning_tokens":2406}},"tokens_in":535,"tokens_out":2486,"duration_ms":15848,"temperature":1.0,"reasoning_tokens":2406,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T12:51:53.456919+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the last-session Macro-F1 on NHIE and GRABMyo when the finetuning regularization uses real stored samples instead of inverted ones, a real-replay upper bound; if AnchorInv falls clearly below that bound on either dataset, the claim that inverted samples carry as much prior-class information as real samples is contradicted. A second check: replace the anchor-guided mean-absolute-error inversion loss with a label-space cross-entropy inversion loss; if performance does not drop, feature-space guidance specifically is not the cause of the reported gains.","supporting_citations":[{"cited_title":"u ller-Putz, G.; Schl \\","cited_arxiv_id":null,"evidence_quote":"BCI Competition IV 2a dataset: one of the three evaluation benchmarks."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"TEEN: a second state-of-the-art vision FSCIL baseline used in the comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"DeepDream: inversion-based generation approach that AnchorInv extends and uses as a comparison baseline."},{"cited_title":"M.; Mallya, A.; Hoiem, D.; Jha, N","cited_arxiv_id":null,"evidence_quote":"DeepInv: establishes model inversion for data-free knowledge transfer; AnchorInv adapts this idea to feature-space targets."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"NC-FSCIL: a state-of-the-art vision FSCIL baseline re-implemented for physiological data."}],"review_version":1}