{"id":"1b330dc9-1b43-42a0-aa56-04e9356967fd","arxiv_id":"2506.09347","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A continual-learning plugin that identifies low-density feature samples as noisy-label errors and erases them by fine-tuning a pseudo-class neuron, improving accuracy and reducing forgetting in most tested settings.","lead":"ErrorEraser is a plugin that tries to detect and erase wrong memories caused by noisy labels while a model learns tasks one after another. It combines density-based outlier detection with a pseudo-class fine-tuning step, and tests on several continual learning methods.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claim that ErrorEraser accurately identifies noisy labels and erases only erroneous knowledge is untested; reported gains may stem from the L_ge loss or from pseudo-class fine-tuning rather than from accurate unlearning.","rationale":"The reader identified the weakest assumption as the reliable detection of mislabeled samples as low-density outliers and the claim that only erroneous knowledge is erased. My analysis agrees and sharpens it: the missing ground-truth validation is not just a reporting gap—it is the load-bearing hinge of the paper's mechanism. The error-identification module has two coupled effects: the L_ge loss already reshapes training by downweighting low-density samples, and the subsequent selection by Eq. (11) uses the same density model on features produced by a backbone trained with that loss. This circularity means that even if selection is poor, the pipeline could still improve performance through the robust-training effect of L_ge and through the fine-tuning/pruning operation acting as a form of regularization on a small subset. The paper's ablation (Table 6) shows both modules contribute, but it does not isolate the contribution of accurate selection. A random-selection baseline would directly test whether the identification step is load-bearing. Additionally, the headline claim of 'lower forgetting rates across three types of CL methods' is contradicted by Table 2 for HAT+ErrorEraser on MNIST (F rises from 1.29 to 7.34) and on CIFAR-10 (F rises from 1.34 to 33.23), which the paper itself acknowledges; this weakens the universality claim but is secondary to the identification issue. Consequently, the appropriate verdict remains CONDITIONAL, requiring the synthetic-noise precision/recall measurement and selection-rule baselines. My read does not change the reader's verdict.","tokens_in":25936,"tokens_out":3304,"duration_ms":38594,"concrete_test":"On CIFAR-10 with 50% asymmetric noise, compute per-task precision and recall of D_s from Eq. (11) against the known noisy-label set. Then run the full ErrorEraser pipeline with three selection rules: density-based D_s, random subset of the same size, and highest-loss subset of the same size. If the S and F gains over the noisy baseline are equivalent across all three, the identification module is not load-bearing. If density-based selection significantly outperforms random and loss-based selection, the identification claim is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central mechanism has two linked premises: (1) noisy-label samples are detectable as low-density outliers in the feature space learned by the model (Sect. 4.1.3, Eq. 2), and (2) fine-tuning only the selected low-density set D_s into a pseudo-class and then pruning the new neuron erases only erroneous knowledge while preserving correct knowledge (Sect. 4.2.2). Neither premise is validated against ground truth. The paper never reports precision or recall of D_s relative to the known synthetic noisy labels, despite using asymmetric noise with known mappings (Sect. 7.1). This matters because the training loss L_ge (Eq. 10) already downweights low-density samples during backbone training, so the post-hoc density-based selection is partially circular: the model is trained to make low-density samples less influential, and then those same samples are flagged as noisy. The reported end-task gains are therefore consistent with a weaker mechanism: L_ge acting as a robust-loss regularizer, plus fine-tuning a small high-loss subset to a pseudo-class, which can improve generalization regardless of whether the selected samples are actually mislabeled. The paper's own sensitivity analysis (Fig. 6) shows performance degrades as the forgotten proportion increases, acknowledging that normal samples get included, but it does not quantify how many are included or whether the gains survive with random or loss-based selection. Without measuring identification accuracy or comparing selection rules, the load-bearing claim that ErrorEraser erases erroneous knowledge rather than merely regularizing the model is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ErrorEraser, a plugin for continual learning (CL) that identifies and erases 'erroneous memories' caused by noisy labels. It has two modules: Error Identification, which uses normalizing flows to model the probability density of task features in a compact latent space, with a new loss L_ge that downweights low-density samples during backbone training; and Error Erasure, which selects the lowest-density samples D_s, fine-tunes them into a pseudo-class, and prunes the corresponding neuron. The authors claim that this removes erroneous knowledge and improves both retention and transfer across three types of CL methods (regularization-, replay-, and optimization-based). Experiments on synthetic noisy MNIST, CIFAR-10, CIFAR-100 and real WebVision report higher composite scores S and lower forgetting for most combinations, with comparisons to noisy-label CL methods and unlearning baselines.","tokens_in":26187,"tokens_out":3268,"duration_ms":38404,"significance":"If the central mechanism is correct, the paper would make a useful contribution: it reframes CL as needing not only to prevent forgetting but also to intentionally forget incorrect knowledge, and it proposes a plug-and-play method that requires no access to historical data. The empirical scope is broad (six CL backbones, three synthetic datasets, one real dataset, several noise rates and task counts), and the code is released. The resource-cost analysis (Table 5) further indicates that the plugin adds modest overhead. However, the significance is tempered by the fact that the load-bearing claim---that low-density samples are noisy and that fine-tuning them into a pseudo-class erases only erroneous knowledge---is never validated against ground-truth noisy labels. The reported gains are also consistent with a weaker mechanism in which L_ge acts as a robust-loss regularizer and the forgetting fine-tuning provides additional regularization on a high-loss subset. The paper therefore needs a direct evaluation of identification accuracy and a more careful ablation of the selection rule before its main claim can be accepted.","major_comments":[{"comment":"The core premise is that samples with low probability density p(x) in the NF latent space are 'potential noisy labels'. Since the synthetic noise is generated with known asymmetric mappings (Sect. 7.1), the authors can directly measure the precision/recall or overlap of the selected set D_s with the true mislabeled set. No such metric is reported. This matters because the training loss L_ge (Eq. 10) already downweights low-density features during backbone training, so the post-hoc density-based selection is partially self-referential: the model is trained to make low-density samples less influential, and those same samples are then flagged as noisy. The concrete test is to report identification accuracy on the known noise labels, and to compare the end-task performance of density-based selection with random selection and with selection based on large classification loss. Without that experiment, the paper's explanatory narrative ('accurately identify and erase erroneous memories') is unsupported, and the gains could be attributed to L_ge plus fine-tuning a high-loss subset as a robust-loss regularizer.","section":"Sect. 4.1.3, Eq. (11)"},{"comment":"The abstract and Section 5.2 claim 'higher accuracy and lower forgetting rates across three types of CL methods'. The appendix data do not fully support this universal claim: on CIFAR-10 at 10% noise, DERPP+ErrorEraser has S=61.81 versus S=63.68 for the baseline DERPP, and at 30% noise it has S=57.00 versus 58.08. Similarly, HAT+ErrorEraser on CIFAR-10 shows substantially higher forgetting (F=33.23 at 50% noise) than the baseline HAT (F=1.34). The paper gives a qualitative explanation for the HAT case, but no analysis for DERPP. Moreover, the experiments are run five times and averaged, yet no standard deviations, confidence intervals, or significance tests are reported anywhere. The claim of 'significant mitigation' (Sect. 5.3, Table 4) is therefore not statistically grounded, and the universal-improvement claim is contradicted by the appendix results without discussion.","section":"Sect. 5.2, Table 2 and Appendix Table 8"},{"comment":"The erasure mechanism assumes that fine-tuning only D_s into a pseudo-class and pruning the new neuron erases erroneous knowledge while leaving correct knowledge intact. The sensitivity analysis (Fig. 6) acknowledges that increasing the forgotten proportion degrades performance because 'normal samples' are included, but it does not quantify how many correct samples are affected, nor does it report the performance of the model on the actual noisy-label subset. Since D_s is defined only by a density threshold with no validation of its label-correctness composition, the safety of the unlearning step is unquantified. A direct measurement of the false-positive rate of D_s (correct samples selected) would settle whether the procedure is truly selective or simply a high-loss fine-tuning step that improves generalization regardless of label correctness.","section":"Sect. 4.2.2, Eq. (13) and Fig. 6"}],"minor_comments":[{"comment":"There are frequent typos and formatting artifacts: 'Taks' in Tables 2, 8, and 9; concatenated numbers in Table 8 ('14.5157.90', '9.0661.12', etc.); and inconsistent spacing around equations. These should be cleaned up.","section":"Throughout"},{"comment":"The composite metric S is introduced without a motivation or a reference to a previously established metric. The formula as printed is S = A1*A2*n / (A1 + A2*n + F_k); please confirm that the intended denominator is indeed A1 + (A2 * n) + F_k and clarify why n multiplies A2 but not A1.","section":"Eq. (16)"},{"comment":"The description of WebVision says 'the top 14 largest classes' and 'seven tasks with randomly paired classes', but the main-text Table 3 shows results for WebVision without specifying the task count. Please align these details and state the noise rate and class/task split consistently.","section":"Sect. 7.1"},{"comment":"In Table 5, the resource numbers for 'Cover' and the baselines are reported with order-of-magnitude values ('~4 x 10^1' seconds). It would be clearer to report exact numbers and to state which hardware and implementation were used, since the efficiency comparison is used to support the plugin's practicality.","section":"Sect. 5.3, Table 5"}],"recommendation":"major_revision","confidential_remarks":"The paper's empirical core is broad and the proposed plugin appears to help in several settings, but the central mechanistic claim is not tested against ground-truth noise labels. I would encourage the editor to request a revision that adds (1) an evaluation of identification precision/recall on the synthetic noise, (2) an ablation comparing density-based selection to random or loss-based selection, and (3) standard deviations or significance tests for the main tables. If the authors are unable to provide these, the paper should be reconsidered, since the title and abstract make a strong claim about accurate identification and erasure."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The interesting contribution is the combination of density-based outlier identification with pseudo-class boundary-shift unlearning for unknown noisy labels in old CL tasks. The end-task gains are broadly reported and mostly positive, but the paper never shows that the flagged samples are actually the mislabeled ones. The gains are consistent with L_ge acting as a robust-loss regularizer, so the unlearning claim is not yet supported.\n\nWhat is new: existing noisy-label CL methods handle noise in the new task or need a known forget set. ErrorEraser targets old-task noise without historical access, using a normalizing flow on the feature space and pruning a pseudo-class neuron. That combination is new. The empirical scope is decent: three CL families, multiple noise rates and task splits, WebVision, ablations, and resource-cost comparisons. The code link is a plus.\n\nWhere it is soft: no standard deviations despite five runs. The identification module is never validated against the synthetic noise ground truth; precision and recall of D_s are not reported. Without that, we do not know if the method removes errors or rare-but-correct samples. The circularity concern is real: the L_ge loss in Eq. (10) already downweights low-density features during training, so the post-hoc density-based selection is partly a reflection of the training loss, not an independent signal. The appendix undermines the abstract's 'across three types' claim: DERPP+ErrorEraser at 10% and 30% noise on CIFAR-10 is worse than the DERPP baseline. The sensitivity analysis shows performance degrades as the forgotten proportion increases, which admits normal samples get included, but it does not quantify how many. A random-selection baseline for D_s is missing.\n\nBottom line: this is a practical, subfield-level contribution that deserves referee time. The authors should add identification metrics, standard deviations, and a random-selection comparison before the unlearning claim can be taken at face value. I would cite it for the plugin's practical gains, not for the mechanism as stated.","headline":"A useful plugin for noisy labels in continual learning with a plausible mechanism, but the identification step is unvalidated and the gains may be partly regularization.","tokens_in":26778,"tokens_out":3044,"would_cite":true,"duration_ms":33733,"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":"Continual learning can be improved by deliberately forgetting memories caused by mislabeled samples, not just by preventing forgetting.","keywords":["continual learning","machine unlearning","noisy labels","data bias","normalizing flows","selective forgetting","catastrophic forgetting","knowledge transfer"],"falsifier":"On a benchmark with known injected noise, compute the precision and recall of the flagging rule p(x) < delta against the ground-truth noisy labels, and compare the effect of removing the flagged set versus removing the same number of random samples from clean data. If flagged sets overlap the true noisy labels no better than chance, or if pruning them hurts clean-data accuracy as much as random pruning, then the reported improvements are regularization effects rather than targeted unlearning.","tokens_in":25703,"feed_emoji":"🧹","tokens_out":6438,"duration_ms":57259,"temperature":0.7,"pith_summary":"Continual learning models trained on streams with noisy labels absorb erroneous knowledge, and this paper argues that those errors are not a side effect but a cause of both worse retention and worse transfer to later tasks. The authors propose that a continual learner should deliberately forget such errors, and they build a plugin, ErrorEraser, that does so without being told which labels are wrong and without revisiting old-task data. ErrorEraser first fits a normalizing-flow density model to the intermediate feature space of the current task, flags the low-density tail as potentially mislabeled, then fine-tunes only those samples into a new pseudo-class and prunes that pseudo-class away. The claim is that this targeted forgetting raises accuracy and lowers forgetting across regularization-, replay-, and optimization-based CL methods, often matching or exceeding clean-data performance. A reader should care because real-world data streams contain noisy labels and errors learned early can amplify through the task sequence.","feed_headline":"Plugin erases noisy-label memories to boost continual learning","feed_subtitle":"Low-density outliers go to a pseudo-class that is then pruned, lifting accuracy and cutting forgetting across CL methods.","key_machinery":"The load-bearing mechanism is a normalizing-flow density model attached to the intermediate feature layer h_b of the continual-learning backbone, trained jointly with the classifier through the combined loss L = L_ce + L_ge + L_cl. The flow maps features to a multivariate normal latent distribution and gives each sample a probability density p(x); the low-density subset D_s = {x_i : p(x_i) < delta} is the flagged erroneous memory. Erasure is done by appending one pseudo-class neuron, fine-tuning the model on D_s with a pseudo-label so those samples' decision space moves into the new neuron, and then pruning that neuron and its associated activations, restoring the original output width while satisfying F_{theta'_t}(x_i) approximately 0 for the flagged samples. This machinery is what makes the claim 'forget unknown errors without historical data' operative.","core_discovery":"On the paper's own terms, the central discovery is that unknown noisy-label bias in old tasks can be identified and unlearned in a continual learner without any access to historical data. The identification uses the probability density of task features in a compact layer of the network, learned by a normalizing flow with a composite loss that couples density learning to the classification loss; samples whose density falls below a threshold delta are treated as representative erroneous samples. The erasure then expands the output layer with one pseudo-class neuron, fine-tunes the model so those low-density samples map to that neuron, and prunes the neuron, approximately zeroing the model's output on the erroneous samples while leaving other samples' activations essentially unchanged. Across CIFAR-10, CIFAR-100, MNIST, and WebVision with injected or real label noise, the authors report that adding ErrorEraser to EWC, MAS, LWF, DERPP, TAT, and HAT improves the harmonic-mean score S, average new-task accuracy A1, average historical accuracy A2, and lowers the forgetting rate F, with some noisy-data results surpassing clean-data baselines.","pith_inferences":["The paper never reports how many flagged samples actually match the injected noisy labels; a precision-and-recall analysis against ground-truth noise would settle whether the gains come from accurate identification or from generic regularization of low-density points.","If low density is a reliable corruption signal, the same density estimate could be used earlier, as a sample-weighting filter during training, potentially making the post-hoc pruning step unnecessary for some noise rates.","Because the method is demonstrated only in task-incremental settings with known task IDs and a multi-head classifier, a natural next test is whether a shared density model can identify mislabeled samples when task boundaries are unknown or classes overlap.","The pseudo-class-and-prune operation is a general selective-forgetting primitive; it may extend to removing poisoned, outdated, or privacy-sensitive instances from a trained model without full retraining."],"forward_implications":["Adding ErrorEraser to any of the six tested CL methods raises the combined retention-and-transfer score S and lowers the forgetting rate F under 50% label noise.","The plugin is agnostic to the CL objective: it works with regularization losses, replay losses, and parameter-isolation masks, so the same unlearning module can wrap a variety of continual learners.","Because only the low-density tail D_s is fine-tuned and then pruned, the erasure is cheap in time, memory, and model size compared with retraining or contrastive purification.","In several settings the noisy-data model with ErrorEraser outperforms the same CL method trained on clean data, which the authors attribute to removal of conflicting or irrelevant knowledge.","The incremental feature-distribution learning strategy lets the density model be updated across tasks instead of retrained from scratch, keeping downstream-task overhead low."],"supporting_citations":[{"why":"supplies the normalizing-flow density model used to estimate p(x) and rank samples by probability density.","marker":"[34]"},{"why":"defines the selective-forgetting CL baseline LSFM and the harmonic-mean metric S that the paper adopts for evaluation.","marker":"[39]"},{"why":"provides the self-purified replay baseline SPR for noisy data streams, a key comparison for generalizability.","marker":"[16]"},{"why":"provides the continual noisy-label baseline CNLL that ErrorEraser is compared against.","marker":"[14]"},{"why":"provides the EWC regularization-based CL method that ErrorEraser is plugged into and ablated with.","marker":"[19]"},{"why":"provides the HAT optimization-based CL method whose rigid masks interact with the plugin in the experiments.","marker":"[37]"},{"why":"provides the DERPP replay-based CL baseline used in the performance and attention-map comparisons.","marker":"[5]"}],"fun_headline_variants":["Erase biased memories to improve continual learning","Unlearn data bias with a universal CL plugin","Prune spurious correlations to boost CL accuracy","Forget erroneous labels to learn tasks better","ErrorEraser plugin removes bias to lift CL performance"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that mislabeled samples are exactly the low-density outliers in the learned feature space, so that pruning the pseudo-class built from those outliers erases only erroneous knowledge and does not discard rare but correct samples.","fun_headline_variants_meta":{"raw":{"variants":["Erase biased memories to improve continual learning","Unlearn data bias with a universal CL plugin","Prune spurious correlations to boost CL accuracy","Forget erroneous labels to learn tasks better","ErrorEraser plugin removes bias to lift CL performance"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000249,"raw_usage":{"total_tokens":1581,"prompt_tokens":1005,"completion_tokens":576,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":621,"completion_tokens_details":{"reasoning_tokens":506}},"tokens_in":621,"tokens_out":576,"duration_ms":6501,"temperature":1.0,"reasoning_tokens":506,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:51:02.479875+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a benchmark with known injected noise, compute the precision and recall of the flagging rule p(x) < delta against the ground-truth noisy labels, and compare the effect of removing the flagged set versus removing the same number of random samples from clean data. If flagged sets overlap the true noisy labels no better than chance, or if pruning them hurts clean-data accuracy as much as random pruning, then the reported improvements are regularization effects rather than targeted unlearning.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the normalizing-flow density model used to estimate p(x) and rank samples by probability density."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"defines the selective-forgetting CL baseline LSFM and the harmonic-mean metric S that the paper adopts for evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"provides the self-purified replay baseline SPR for noisy data streams, a key comparison for generalizability."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"provides the continual noisy-label baseline CNLL that ErrorEraser is compared against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"provides the EWC regularization-based CL method that ErrorEraser is plugged into and ablated with."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"provides the HAT optimization-based CL method whose rigid masks interact with the plugin in the experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"provides the DERPP replay-based CL baseline used in the performance and attention-map comparisons."}],"review_version":1}