{"id":"b8629b50-f1ed-4231-9248-78e26c4c1df1","arxiv_id":"2411.13886","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"CLFace removes the classification layer from continual face recognition and uses multiscale, geometry, and contrastive distillation to update a student model without labels or exemplars, reporting gains on unseen-identity benchmarks.","lead":"CLFace updates face recognition models on streams of new identities without storing old face images or using labels, by distilling knowledge from a frozen teacher into a student model. The paper reports small but consistent gains over prior continual learning methods on unseen identities, but never measures whether previously learned identities are forgotten.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper never evaluates accuracy on identities from previous incremental steps, so its claim of preserving learned knowledge is unsupported; the reported gains on unseen identities could hide severe forgetting.","rationale":"The reader's identified weakest assumption matches the most load-bearing concern: the evaluation protocol never tests identities learned in earlier incremental steps. The paper's abstract and introduction claim that CLFace preserves and incrementally extends learned knowledge, but Section 3.1 explicitly defines the test set as disjoint from all training identities, and Section 4.3 reports only on that disjoint set. This means the central continual-learning property, retention of previously learned identities, is not measured. The distillation losses are applied only on the current batch of new data; they align student and teacher outputs for those new identities, but provide no constraint on the student's outputs for old identities. Since the model does not store exemplars and no old data is revisited, there is no mechanism in the training loop that is even shown to preserve old identity geometry. The proposed check, evaluating final-model accuracy on held-out identities from each incremental step and comparing to the step-specific teacher, would directly settle whether forgetting occurs. If the check shows no significant drop, the paper's claim is supported; if it shows a drop, the central claim fails. This concern is addressable with an additional evaluation, so the reader's CONDITIONAL verdict remains appropriate. The paper's contribution is plausible, but the missing measurement is essential to the stated goal. I agree with the reader's assessment that this is the core weakness and that additional evidence is needed before the claim can be accepted.","tokens_in":14708,"tokens_out":3840,"duration_ms":39943,"concrete_test":"After the final incremental step T, build a held-out verification set from each earlier step t by splitting D(t) into disjoint training and held-out identity sets. Run 1:1 verification (e.g., TAR@FAR=1e-4) on pairs from the held-out identities of step t using the final model, and compare with the same pairs evaluated by the model from step t (the teacher at that time). If the final model's accuracy on early-step identities drops by more than 1% absolute, then CLFace has not preserved previously learned knowledge.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing concern is that the experimental protocol cannot detect catastrophic forgetting of previously learned identities. Section 3.1 defines the in-domain test set D_test as having identities disjoint from all training steps (C_t ∩ C_test = ∅), and Section 4.3 reports verification accuracy only on that disjoint set. The paper's stated goal is to preserve and incrementally extend the learned knowledge (Abstract, Section 1), and its own design criteria (Section 1, properties I–IV) require retaining past knowledge without storing old data. But no experiment measures accuracy on identities introduced in earlier incremental steps, so a model that completely forgets C(1)...C(T-1) would still obtain the reported scores. The distillation losses (MSFD, GPKD, CKD) are computed only on the current step's data; they align student and teacher outputs on new identities but impose no constraint on the student's outputs for old identities. Since no old data or exemplars are retained, there is no mechanism that is even tested for preserving old identity geometry. Therefore the central claim of lifelong preservation is unverified, and the reported improvements over baselines on unseen identities could coexist with complete forgetting of earlier enrollments.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CLFace, a continual learning framework for face recognition that removes the classification layer and avoids storing exemplars. A frozen teacher model supervises a student model through three distillation losses: multiscale feature distillation (MSFD), geometry-preserving knowledge distillation (GPKD), and contrastive knowledge distillation (CKD). The model is trained incrementally on batches of new face identities and evaluated on unseen identities from in-domain datasets (VGGFace2, MS1MV2, WebFace12M) and out-of-domain benchmarks (LFW, AGEDB, CALFW, CFPFP, CPLFW, IJB-B, IJB-C, TinyFace). The paper reports that CLFace outperforms fine-tuning, LwM, PODNet, and CRL, and provides ablations of the loss components and the amount of base training data.","tokens_in":14919,"tokens_out":7684,"duration_ms":73337,"significance":"If the central claim were fully validated, the framework would be practically significant: it is label-free during incremental steps, uses a fixed-size architecture, and does not store old samples, which is attractive for privacy-constrained biometric deployments. The experimental effort is substantial, including large-scale datasets, multiple incremental-step settings, out-of-domain benchmarks, and ablation studies. However, the significance is conditional because the protocol never measures accuracy on identities learned in earlier incremental steps, so the paper's stated goal of preventing catastrophic forgetting of previously learned identities is not directly supported.","major_comments":[{"comment":"The central claim that CLFace preserves learned knowledge and mitigates catastrophic forgetting is never directly measured. In §3.1 the in-domain test set D_test is defined with identities disjoint from all training steps (C_t ∩ C_test = ∅ for t = 1,...,T), and the results in Figure 3 and Tables 2–5 are all verification accuracies on this disjoint set. A model that completely forgets the identities from steps 1,...,T−1 would still obtain these scores, because the total loss in Eq. (5) is computed only on the current step's batch and no exemplars or prototypes are retained. The inference in §4.5 that incremental improvement on unseen identities 'confirms that CLFace effectively mitigates CF' is therefore not valid. I recommend adding a direct backward-transfer evaluation, e.g., reporting verification or closed-set identification accuracy on the identities from each previous step after each incremental step, or, if the authors intend to claim only open-set generalization, removing the catastrophic-forgetting claims from the Abstract and Section 1.","section":"§3.1, §4.3–4.4"},{"comment":"The ablation that is supposed to show the effect of removing identity supervision does not display the comparison it describes. Table 5's final row is 'ID + MSFD + GPKD + CKD', but there is no row for 'MSFD + GPKD + CKD' without ID; the text compares the no-ID objective from Table 3 against 'ID + GPKD' from Table 5. Since the reported conclusion is that 'replacing ID supervision with CKD loss does not compromise generalizability', the table should include the exact no-ID row, or the text should explicitly state that the comparison is across Tables 3 and 5.","section":"§4.5, Table 5"},{"comment":"The reported improvements over the strongest baseline CRL are very small—between 0.0 and 0.35 percentage points in Table 2 and around 0.04–0.10 points in Figure 3—yet the paper gives only the mean over three random class orders with no standard deviation or significance measure. Given that the claimed advantage is of the same order as typical run-to-run variation in face-verification benchmarks, the comparison is not robust without error bars. Please report mean ± std and, if possible, the per-order values.","section":"§4.2, Table 2, Figure 3"}],"minor_comments":[{"comment":"The algorithm title reads 'CL2Face Implementation' and should be 'CLFace Implementation'.","section":"Algorithm 1 caption"},{"comment":"The label 'Contrastiave Knowledge Distillation' contains a typo; it should be 'Contrastive'.","section":"Figure 2"},{"comment":"The acronym for the multiscale feature distillation loss is inconsistent: 'MFSD' appears in the contributions bullet list, 'MSD' appears in the ablation text, and Eq. (2) defines 'MSFD'. Use one spelling throughout.","section":"§3.3, §4.5"},{"comment":"The formulation allows overlapping identities across steps ({C_t ∩ C_t'} ≠ ∅), but the experimental protocol in §4.2 uses disjoint sets. Clarify whether the overlap scenario is ever tested, since it is one of the properties claimed in Section 1.","section":"§3.1"},{"comment":"The mixed-quality and low-quality columns for LWF-MC, LwM, PODNet, and CRL are all dashes, which is inconsistent with the text's claim that CLFace outperforms these methods 'across all datasets'. Complete these entries or restrict the claim to the datasets where comparisons exist.","section":"Table 2"},{"comment":"The phrase 'open-set face recognition' is not standard: the protocol evaluates on unseen identities but does not include an open-set rejection (unknown-class) component. Please qualify the terminology.","section":"§1, §4.6"},{"comment":"The claim of 'label-free supervision' applies only to incremental steps; the base model is trained with ArcFace label supervision. Please state this explicitly.","section":"§2, §4.2"},{"comment":"In the discussion of the base-training ablation, the text says 'Rank-1' for TinyFace while Table 4 and Table 2 consistently use Rank-5. Correct this inconsistency.","section":"§4.5, Table 4"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is readable and the experimental effort is substantial, but the missing seen-identity evaluation is a substantive gap. I would support publication after the authors either add a direct forgetting measurement or significantly narrow the claims. I also note that no code is provided, which makes reproducing the exact numbers harder, although the implementation details are reasonably complete."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new piece here is the packaging: removing the classification layer during incremental steps, training a student initialized from a frozen teacher with three distillation losses (multiscale feature distillation, geometry-preserving cosine distillation, InfoNCE-style contrastive distillation), and evaluating on identities disjoint from all training steps. That is a reasonable recipe for a privacy-constrained, exemplar-free biometric update, and the paper is honest that the losses are borrowed from existing CL work. The experiments are large-scale and the ablation studies are useful; the gains over CRL and fine-tuning are small but consistent on out-of-domain benchmarks.\n\nThe soft spot is load-bearing. The paper's stated goal is to preserve and incrementally extend learned knowledge, but no experiment measures accuracy on identities introduced in earlier incremental steps. The test sets are explicitly disjoint from all training identities (Section 3.1, Figure 1), so a model that completely forgets all prior enrollments would still produce the reported scores. The distillation losses are computed only on current-step data; they align student and teacher on new identities but impose no direct constraint on outputs for old identities. Without exemplars or prototypes, there is no mechanism even tested for retaining old identity geometry. The paper's own Table 4 shows small improvements over steps on unseen identities, but that is not evidence against forgetting of seen identities. The stress-test note is right: this is the central claim, and it is unverified.\n\nOther issues are more minor. Reported gains are tiny (0.04–0.35%), no error bars or significance tests appear, and no code is provided. Hyperparameters are chosen by grid search on the same benchmarks used for final evaluation, so there is some selection risk, though not enough to sink the paper on its own.\n\nWho is this for? Researchers working on continual learning for face recognition or biometrics, especially those interested in label-free, exemplar-free update schemes. They would get a clear, well-written baseline method and a cautionary example of an evaluation protocol that sidesteps the forgetting question. The paper deserves a serious referee: the idea is sensible, the experiments are substantial, and the flaw is addressable by adding a seen-identity evaluation (e.g., verification on identities from earlier steps, or a probe set mixing old and new). I would not cite it yet, but I would engage with it.\n\nRecommendation: send to peer review with a request for major revision adding direct forgetting measurements on previously learned identities and, ideally, error bars and code.","headline":"Plausible distillation recipe for label-free incremental face recognition, but the paper never measures accuracy on previously learned identities, so its central anti-forgetting claim is unsupported.","tokens_in":15432,"tokens_out":704,"would_cite":false,"duration_ms":9044,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a face recognition model can keep learning new identities from a continuous stream without storing old images or using identity labels, by training a student on three distillation losses from a frozen teacher.","keywords":["continual learning","face recognition","knowledge distillation","catastrophic forgetting","open-set face recognition","label-free supervision","biometric identification","incremental learning"],"falsifier":"Run the same incremental protocol but, after every step, test on held-out samples of the identities trained in every previous step, as in standard class-incremental evaluation. If verification or identification accuracy on those old identities drops substantially after later steps, the distillation objective has not actually preserved the learned knowledge, even though unseen-identity accuracy could remain high.","tokens_in":14514,"feed_emoji":"👤","tokens_out":8677,"duration_ms":79210,"temperature":0.7,"pith_summary":"The paper claims that a face recognition model can keep learning new identities from a continuous stream without storing old images and without identity labels, by having a frozen teacher model supervise a student model through three distillation losses. If true, this gives a privacy-friendly and resource-light way to update deployed biometric systems without retraining from scratch or retaining sensitive face data. The authors report that this classification-free, label-free objective improves verification accuracy on unseen identities, both from the same dataset and from unrelated benchmark datasets, compared with fine-tuning and with established continual learning baselines. The paper's central assertion is that a fixed architecture trained only on distillation can preserve old knowledge and keep improving on new identities at the same time.","feed_headline":"Distilled face model learns new identities without old data","feed_subtitle":"CLFace learns new faces on a stream, with no stored exemplars and no labels, while keeping accuracy on unseen faces.","key_machinery":"The carrying mechanism is teacher-student distillation with a frozen teacher and a student initialized from it. MSFD transfers local spatial structure: for each intermediate stage except the first, both models' feature maps are pooled across channels and L2-normalized, and the squared L2 distance between the resulting attention maps is penalized. GPKD transfers global geometry: cosine similarity between the teacher's and student's normalized embeddings is maximized so the embedding space is not rotated. CKD adds discriminative contrast: within each batch, the positive pair is the teacher and student embedding of the same identity, and all other identities are negatives, optimized through an InfoNCE-style loss. Removing the classification layer keeps the architecture and parameter count fixed across steps, which is what makes the approach scalable and label-free.","core_discovery":"CLFace is a continual learning framework for face recognition in which a student model, initialized from a frozen teacher, is trained on each new batch of identities with no classification layer and no label supervision. The objective is a weighted sum of three losses: multiscale feature distillation (MSFD), which takes channel-wise pooled and L2-normalized spatial attention maps at intermediate stages and minimizes their L2 distance between student and teacher; geometry-preserving knowledge distillation (GPKD), which keeps the cosine orientation of the L2-normalized embeddings aligned; and contrastive knowledge distillation (CKD), an InfoNCE-style loss that pulls the teacher and student embeddings of the same identity together while pushing embeddings of different identities apart. The paper's claim is that this combination lets the student inherit the teacher's learned representation, avoids catastrophic forgetting, and sharpens discrimination of new identities. In experiments with 5-step and 10-step protocols on VGGFace2, MS1MV2, and WebFace12M, CLFace achieves higher verification accuracy on unseen identities than fine-tuning, LwM, PODNet, and CRL, on both in-domain and out-of-domain benchmarks including LFW, AGE-DB, CALFW, CFP-FP, CPLFW, IJB-B, IJB-C, and TinyFace.","pith_inferences":["Inference: because the evaluation only covers unseen identities, the reported gains may reflect improved general feature quality rather than actual retention of identities from earlier steps; an evaluation on old identities would separate the two effects.","Inference: the ablation results suggest the three losses play distinct roles—MSFD preserves spatial structure, GPKD stabilizes embedding orientation, and CKD adds contrastive discrimination among new identities—so the same recipe may transfer to other embedding-based incremental metric-learning tasks such as person re-identification or speaker verification.","Inference: the reported improvements over CRL are small on high-quality benchmarks, which hints that the main advantage may be architectural, namely removing the classification layer and label supervision, rather than the specific choice of distillation losses; isolating this by ablating CRL's own losses without its classifier would test that."],"forward_implications":["Deployed face recognition systems could be updated continuously on streams of new identities without storing exemplars, removing a major privacy and storage bottleneck.","Because the architecture is fixed and classifier-free, memory and compute use stay bounded as the number of enrolled identities grows.","The label-free objective permits incremental training on unlabeled face collections, which supports open-set recognition in deployment settings.","The reported pattern that larger incremental steps produce larger gains implies the benefit of lifelong updates grows with the volume of the identity stream."],"supporting_citations":[{"why":"Supplies the ArcFace loss, the iResNet50 backbone, and the MS1MV2 dataset used to pre-train the base model and to evaluate in-domain performance.","marker":"[7]"},{"why":"Introduces learning without forgetting, the distillation-based continual learning idea that CLFace extends and uses as a comparison baseline.","marker":"[20]"},{"why":"Provides the attention-distillation baseline LwM that CLFace is compared against in the in-domain and out-of-domain experiments.","marker":"[8]"},{"why":"Supplies the pooled-output distillation approach across multiple stages that MSFD adapts and that serves as the PODNet baseline.","marker":"[9]"},{"why":"Contributes the embedding-orientation distillation idea that GPKD builds on to preserve the geometry of the feature space.","marker":"[14]"},{"why":"Defines the InfoNCE contrastive objective that CKD uses to make new-identity features discriminative without labels.","marker":"[29]"},{"why":"Provides CRL, the continual representation learning method for biometrics that is the primary comparison baseline and the approach CLFace claims to improve on.","marker":"[43]"}],"fun_headline_variants":["Face model learns new people without old data or labels","Lifelong face recognition: no storage, no forgetting","Distilled face AI learns new identities from a live stream","CLFace: continual learning for faces without past data","Open-set face learning that never forgets, stores nothing"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The experiments never evaluate accuracy on identities the model saw in earlier incremental steps, so the claim that CLFace preserves previously learned knowledge rests on the assumption that stable or improved accuracy on unseen identities is enough to prove no catastrophic forgetting occurred.","fun_headline_variants_meta":{"raw":{"variants":["Face model learns new people without old data or labels","Lifelong face recognition: no storage, no forgetting","Distilled face AI learns new identities from a live stream","CLFace: continual learning for faces without past data","Open-set face learning that never forgets, stores nothing"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000709,"raw_usage":{"total_tokens":3246,"prompt_tokens":1054,"completion_tokens":2192,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":670,"completion_tokens_details":{"reasoning_tokens":2113}},"tokens_in":670,"tokens_out":2192,"duration_ms":17048,"temperature":1.0,"reasoning_tokens":2113,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:46:50.115213+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same incremental protocol but, after every step, test on held-out samples of the identities trained in every previous step, as in standard class-incremental evaluation. If verification or identification accuracy on those old identities drops substantially after later steps, the distillation objective has not actually preserved the learned knowledge, even though unseen-identity accuracy could remain high.","supporting_citations":[{"cited_title":"ArcFace: additive angular margin loss for deep face recognition","cited_arxiv_id":null,"evidence_quote":"Supplies the ArcFace loss, the iResNet50 backbone, and the MS1MV2 dataset used to pre-train the base model and to evaluate in-domain performance."},{"cited_title":"Learning without forgetting","cited_arxiv_id":null,"evidence_quote":"Introduces learning without forgetting, the distillation-based continual learning idea that CLFace extends and uses as a comparison baseline."},{"cited_title":"Learning without mem- orizing","cited_arxiv_id":null,"evidence_quote":"Provides the attention-distillation baseline LwM that CLFace is compared against in the in-domain and out-of-domain experiments."},{"cited_title":"PODNet: pooled outputs dis- tillation for small-tasks incremental learning","cited_arxiv_id":null,"evidence_quote":"Supplies the pooled-output distillation approach across multiple stages that MSFD adapts and that serves as the PODNet baseline."},{"cited_title":"Learning a unified classifier incrementally via re- balancing","cited_arxiv_id":null,"evidence_quote":"Contributes the embedding-orientation distillation idea that GPKD builds on to preserve the geometry of the feature space."},{"cited_title":"Continual representation learning for biometric identification","cited_arxiv_id":null,"evidence_quote":"Provides CRL, the continual representation learning method for biometrics that is the primary comparison baseline and the approach CLFace claims to improve on."}],"review_version":1}