{"id":"e94533b5-3040-465b-bb01-4c5d341fa0eb","arxiv_id":"1908.09747","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Multiplying a face-recognition loss by a sigmoid of that loss, which up-weights hard examples and down-weights easy ones, gives small and inconsistent accuracy changes on LFW and YTF.","lead":"This paper reweights face-recognition loss functions with a sigmoid so that hard examples get more weight during training. The authors report small accuracy gains in some settings and small losses in others when applying the trick to Cross-Entropy, SphereFace, and ArcFace.","discovery_kind":"incremental","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (2)/(4) do not specify whether L is per-sample or batch-averaged; under the batch-level reading the reweighting factor is a constant, so no hard-mining occurs.","rationale":"The reader's hyperparameter-sensitivity concern is real, but it presumes the mechanism is correctly specified. The scalar-vs-per-example ambiguity is more load-bearing: if the batch-averaged reading is the implemented one, the proposed loss is just a global scaling of the base loss, so the reported accuracy differences cannot be attributed to hard mining. The text and Fig. 1 strongly suggest per-example intent, so this is probably a presentation/implementation-description gap rather than evidence of misconduct; still, the formula as written is internally inconsistent (Algorithm 1 also multiplies by an extra β). I therefore keep the CONDITIONAL verdict, with the condition expanded to require an unambiguous per-sample definition and preferably code. My read differs from the reader's weakest-assumption choice: hyperparameters are secondary until the loss form is pinned down.","tokens_in":8108,"tokens_out":10705,"duration_ms":110923,"concrete_test":"On a two-sample batch (one easy, p=0.9; one hard, p=0.1), compute the gradient of L_HM_CE exactly as written in Eqs. (1) and (4), and compare the ratio of per-sample gradient norms to that of plain CE. If the ratio is unchanged, the scalar reading cannot produce hard-mining. Then recompute with per-example L_i in L_HM = (α/N) Σ_i L_i σ(β L_i) and confirm the ratio changes. This settles whether the specified loss can possibly explain the reported boosts.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim depends on the multiplier σ(βL) being different for hard and easy examples within a batch. But Eq. (1) defines L_CE as the batch average, and Eq. (4) substitutes that scalar into σ(βL). Under that reading, dL_HM/dθ = α[σ(βL)+βLσ'(βL)] dL_CE/dθ; the bracketed term is identical for every sample, so relative gradients are unchanged. The method then reduces to a global loss scaling and cannot 'increase the loss for harder examples and decrease it for easier examples.' Algorithm 1 adds a second ambiguity: it sets x=β×L and returns α×x×z, i.e., αβL σ(...), not α×L×σ(...) as in Eq. (4). If the authors intend per-example losses, the equations, algorithm, and reduction over the batch must be stated explicitly; as written, the formal definition does not implement the claimed per-sample hard-mining mechanism.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a family of loss functions, called Hard-Mining loss, defined by multiplying an existing loss L by a sigmoid of L: L_HM = α L σ(βL), with σ(x)=1/(1+exp(-A(x-B))). The mechanism is intended to up-weight hard examples and down-weight easy examples during training. The authors apply the wrapper to Cross-Entropy, Angular-Softmax, and ArcFace losses, training ResNet18 on CASIA-WebFace or MS-Celeb-1M and evaluating verification accuracy on LFW and YTF. The central claim, stated in the abstract and conclusion, is that existing loss functions are boosted when used inside the proposed framework. The paper reports improvements in most of the six base-loss/dataset comparisons from Table 1 and Table 2, while acknowledging that some results are comparable.","tokens_in":8256,"tokens_out":4214,"duration_ms":45325,"significance":"If realized correctly, the proposed loss is simple and generic: it is a per-sample weighting scheme that requires no pair selection or external mining procedure and is applicable to any differentiable base loss. That feature is a genuine strength, and Figure 1 gives an intuitive illustration of the intended weighting behavior. However, the contribution is currently not established: the formal equations do not unambiguously define a per-sample reweighting, the algorithm contradicts the equations, the hyperparameters are chosen without sensitivity analysis, and the experimental support consists of single-run accuracy differences that are small and, in three cases, negative. The paper is therefore a promising idea with unsupported claims, and the central mechanism needs to be clarified and re-validated before the proposed loss can be accepted as a reliable contribution.","major_comments":[{"comment":"The formal definition does not implement the claimed per-example mechanism. Equation (1) defines L_CE as a batch-averaged scalar, and Eq. (4) substitutes that scalar into σ(βL_CE). Under that reading, the reweighting factor is identical for every sample in the batch, so the gradient of L_HM_CE is a common multiple of the gradient of L_CE, and the loss cannot \"increase the loss for harder examples and decrease it for easier examples.\" If per-sample losses l_i are intended, the equations must introduce them explicitly, e.g., L_HM_CE = (1/N) Σ_i α l_i σ(β l_i), and the notation must be changed consistently. In addition, Algorithm 1 line 5 returns α × x × z with x = β × L_CE, which equals α β L_CE σ(...), not α L_CE σ(...) as in Eq. (4); the extra factor β makes Algorithm 1 inconsistent with the equation. This ambiguity is load-bearing because the central claim depends on the sigmoid factor varying across samples.","section":"Section 2.2, Eqs. (2) and (4), and Algorithm 1"},{"comment":"The four hyperparameters α=1.5, β=1.1, A=35, and B=0.75 are set empirically with no ablation and no validation split. With A=35 and B=0.75, σ is effectively a near-step function at L≈0.75, so the proposed loss is essentially a hard threshold imposed on the per-sample loss; small changes in B or the loss scale could drastically change which samples are mined. Because the same four values are assumed to work across all three base losses and both training datasets without any sensitivity analysis, the comparisons in Tables 1 and 2 reflect one hand-tuned configuration rather than a validated, generic algorithm. An ablation study or a sensitivity table varying α, β, A, and B is necessary to support the claim that the method is generic.","section":"Section 3.4 and Fig. 1"},{"comment":"The abstract and conclusion claim that existing losses are \"boosted\" by the proposed framework, but the data in Table 2 contradict that in three of six comparisons: LHM_AS on YTF drops from 94.1 to 93.8, LHM_AF on YTF drops from 95.1 to 94.9, and LHM_CE on LFW ties at 95.1. The positive differences elsewhere are at most 0.4 percentage points (e.g., LHM_AF on LFW, 98.0 vs 97.6), and the paper reports no error bars, no repeated runs, and no statistical significance. Consequently, the evidence does not support the strong \"boosted\" claim; at best it shows comparable accuracy in a few configurations. The authors should either soften the claim or provide statistical validation and specify the exact evaluation protocol (e.g., standard 10-fold LFW/YTF verification protocol).","section":"Section 4, Tables 1 and 2"}],"minor_comments":[{"comment":"The text states that \"σ is defined in (1)\", but the sigmoid is defined in Eq. (3), not Eq. (1).","section":"Section 2.3"},{"comment":"There are typos in the table headers: \"Angular-Sofmax\" should be \"Angular-Softmax\" in Table 1 and \"Cross-Entropy los\" should be \"Cross-Entropy loss\" in Table 2.","section":"Tables 1 and 2"},{"comment":"Several minor grammatical and typographical errors should be corrected, including \"intution\" in Section 2.1, \"perfomed\" in Section 4, \"we propose\" capitalized mid-sentence in the abstract, and inconsistent use of \"loss\" vs \"losses\".","section":"Throughout"},{"comment":"The description of the MS-Celeb-1M data is incomplete: the \"cleaned and refined subset\" should be specified precisely (e.g., a citation to the specific cleaned list or a description of the cleaning procedure), since the noise level of that dataset materially affects training outcomes.","section":"Section 3.2"}],"recommendation":"major_revision","confidential_remarks":"The paper currently sits between a workshop-style empirical report and a complete journal contribution. The central idea is simple, but the equations and Algorithm 1 must be reconciled and the per-sample nature of the loss must be stated unambiguously; without that fix, the mechanism is formally only a global scaling of the base loss. The experimental section also needs more than single-run percentages to support the 'boosted' language. If the authors can supply the per-sample formulation, an ablation of the hyperparameters, and a cautious interpretation of the tables, the paper would be within reach of acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this paper reweights a base loss by a sigmoid of the loss itself, which is a close relative of focal loss, but it never cites or compares against focal loss. Worse, the formal definition has a batch-average vs per-sample ambiguity that could mean the method does not actually do what it claims. The experiments show small improvements that are sometimes negative, so the central \"boosting\" claim is overstated. Still, the idea is simple and the numbers are reported transparently.\n\nWhat's genuinely new: the specific choice of a sigmoid on the loss value as a hard-mining multiplier. Focal loss down-weights easy examples via a probability-based term; this uses a sigmoid on the scalar loss. That is a cosmetic difference if the loss is per-sample, and a no-op if it is batch-averaged. The paper does not clarify which.\n\nWhat it does well: the writing is clear, the benchmarks are standard, and the results are not cherry-picked—Table 2 shows two negative results on YTF that the conclusion glosses over but the data still contains. Using three base losses is a reasonable test bed.\n\nSoft spots:\n\n1. The formal ambiguity is load-bearing. Equations (2) and (4) define L as the batch-averaged loss in (1). If that is the intended reading, sigma(beta*L) is a constant for the whole batch, all per-sample gradients are scaled identically, and no hard mining occurs. Algorithm 1 also multiplies by beta, giving alpha*beta*L*sigma(...) rather than alpha*L*sigma(...). If per-sample losses are intended, the equations need to say so explicitly. As written, the formal definition does not implement the claimed mechanism—the stress-test note is correct.\n\n2. No comparison to focal loss, the obvious prior. The cited hard-mining works [2,21,32] are related but none down-weight easy examples through a continuous multiplier on the loss itself. Focal loss is the closest and its omission is a real gap.\n\n3. Empirical support is thin: no error bars, no repeated runs, and the four hyperparameters are picked without ablations or a validation set. The gains on LFW/CASIA are up to 1.4 points for Cross-Entropy, but for Angular-Softmax and ArcFace they are 0.1–0.2 points, and the MS-Celeb-to-YTF results are negative in two of three cases.\n\nWho this is for: a reader working on loss design for face recognition might skim it as a related method. It deserves a serious referee because the formal issue is exactly the kind of thing a review should catch, and the missing focal-loss baseline is fixable. My recommendation: conditional accept, requiring the authors to clarify the per-sample definition, add a focal-loss comparison, and include a sensitivity analysis for alpha, beta, A, and B. Without that clarity, the contribution is hard to distinguish from global loss scaling.","headline":"A minor focal-loss variant with an ambiguous formal definition and thin experiments; worth a conditional referee if the authors can fix the equations and add the missing baseline.","tokens_in":8842,"tokens_out":3828,"would_cite":false,"duration_ms":39273,"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":"Multiplying any face-recognition loss by a sigmoid of its own value shifts training weight to hard examples and improves verification accuracy.","keywords":["face recognition","hard-mining loss","loss function","deep learning","sigmoid weighting","cross-entropy loss","angular softmax","arcface"],"falsifier":"Re-run the CASIA-WebFace to LFW comparison with the same code but set $B$ to 0.5 and again to 1.0, leaving every other setting fixed. If the Hard-Mining losses do not beat their unwrapped baselines under either neighboring $B$ value, the reported boost is tied to the exact hand-set hyperparameters rather than to the reweighting mechanism.","tokens_in":7872,"feed_emoji":"👤","tokens_out":7734,"duration_ms":75064,"temperature":0.7,"pith_summary":"Face-recognition losses are usually dominated by easy examples, which are plentiful, so hard examples contribute little to the gradient. This paper proposes a generic wrapper that multiplies any loss $L$ by $\\sigma(\\beta L)$, a sigmoidal factor that grows with the loss value, so hard examples get up-weighted and easy ones are down-weighted. The paper argues this wrapper is loss-agnostic and demonstrates it by wrapping Cross-Entropy, Angular-Softmax, and ArcFace losses, training ResNet18 on CASIA-WebFace or MS-Celeb-1M, and testing on LFW and YTF. Across the reported settings, the wrapped losses match or exceed their unwrapped baselines in verification accuracy. The underlying claim is that simply reweighting by loss magnitude is enough to make a network learn hard-example characteristics.","feed_headline":"Hard-example reward lifts face recognition across three losses","feed_subtitle":"Multiplying cross-entropy, angular-softmax, or ArcFace loss by a sigmoid boosts LFW and YTF verification accuracy.","key_machinery":"The load-bearing object is the Hard-Mining transform $L \\mapsto \\alpha L \\sigma(\\beta L)$, where $\\sigma$ is a shifted logistic sigmoid with slope and center $A$, $B$. It carries the argument by turning the base loss value itself into a per-sample importance weight, so no explicit hard-sample search or pair selection is needed. Because the factor is monotone in $L$, samples that the current model finds difficult receive larger gradients automatically, and the definition of 'hard' shifts across training as the model improves.","core_discovery":"On the paper's own terms, the central claim is that loss-value-dependent reweighting is an effective and portable mechanism for hard-example emphasis in face recognition. The proposed Hard-Mining loss is defined as $L_{HM} = \\alpha L \\sigma(\\beta L)$, with $\\sigma(x) = (1+e^{-A(x-B)})^{-1}$, so an example with a large base loss is multiplied by a factor near one while an example with a small loss is multiplied by a factor near zero. With $\\alpha=1.5$, $\\beta=1.1$, $A=35$, and $B=0.75$, the wrapper lifts LFW accuracy from 95.35% to 96.75% for Cross-Entropy, from 97.12% to 97.3% for Angular-Softmax, and from 97.79% to 97.9% for ArcFace when trained on CASIA-WebFace; similar or smaller gains are reported on YTF and on MS-Celeb-1M training.","pith_inferences":["An implication the paper leaves implicit: because the wrapper uses only the scalar loss value, the same formula should apply to other classification tasks where easy samples dominate, such as object detection; this is not demonstrated here.","The paper fixes four hyperparameters without an ablation, so a natural extension is to map how the gain varies with the sigmoid center $B$; that map could reveal whether the mechanism or the tuning produces the reported boost.","The sigmoid reweighting is one instance of the broader idea of confidence-based loss scaling, so the reported gains are likely to generalize to other confidence-based losses."],"forward_implications":["The wrapper is loss-agnostic, so the same Hard-Mining formula can be applied to margin-based losses beyond the three tested ones.","The reported accuracy gains, while small, appear on both test datasets and both training sets, so the effect is not tied to one benchmark.","Since the weighting uses only the current loss value, the method needs no separate hard-sample mining stage during training.","In most reported cells the Hard-Mining version is better than or equal to its base loss, supporting the paper's conclusion that easy examples dominate without reweighting."],"supporting_citations":[{"why":"ArcFace loss is one of the base losses the Hard-Mining wrapper is applied to.","marker":"[3]"},{"why":"Angular-Softmax loss from SphereFace is the second base loss used in the framework.","marker":"[17]"},{"why":"ResNet18 provides the training architecture for all experiments.","marker":"[10]"},{"why":"MS-Celeb-1M is one of the two training datasets.","marker":"[9]"},{"why":"CASIA-WebFace is the other training dataset.","marker":"[33]"},{"why":"LFW is the primary testing dataset for verification accuracy.","marker":"[12]"},{"why":"YTF is the video-based testing dataset for verification accuracy.","marker":"[31]"},{"why":"Face alignment preprocessing used before training and testing.","marker":"[34]"}],"fun_headline_variants":["Hard-mining loss boosts face recognition on LFW and YTF","Reweighting loss by hardness lifts face recognition accuracy","Sigmoid-scaled loss improves face recognition across datasets","Hard-example weighting multiplies face recognition accuracy","Portable loss tweak boosts face recognition across losses"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the single hand-chosen set of hyperparameters $\\alpha=1.5$, $\\beta=1.1$, $A=35$, $B=0.75$ works across all three base losses and both training datasets without per-loss tuning.","fun_headline_variants_meta":{"raw":{"variants":["Hard-mining loss boosts face recognition on LFW and YTF","Reweighting loss by hardness lifts face recognition accuracy","Sigmoid-scaled loss improves face recognition across datasets","Hard-example weighting multiplies face recognition accuracy","Portable loss tweak boosts face recognition across losses"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000826,"raw_usage":{"total_tokens":3656,"prompt_tokens":1039,"completion_tokens":2617,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":655,"completion_tokens_details":{"reasoning_tokens":2537}},"tokens_in":655,"tokens_out":2617,"duration_ms":19245,"temperature":1.0,"reasoning_tokens":2537,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:15:33.889791+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the CASIA-WebFace to LFW comparison with the same code but set $B$ to 0.5 and again to 1.0, leaving every other setting fixed. If the Hard-Mining losses do not beat their unwrapped baselines under either neighboring $B$ value, the reported boost is tied to the exact hand-set hyperparameters rather than to the reweighting mechanism.","supporting_citations":[{"cited_title":"In: European Conference on Computer Vision","cited_arxiv_id":null,"evidence_quote":"MS-Celeb-1M is one of the two training datasets."},{"cited_title":"IEEE (2011)","cited_arxiv_id":null,"evidence_quote":"YTF is the video-based testing dataset for verification accuracy."},{"cited_title":"IEEE Signal Processing Letters23(10), 1499–1503 (2016)","cited_arxiv_id":null,"evidence_quote":"Face alignment preprocessing used before training and testing."}],"review_version":1}