{"id":"45d8c559-f0d6-4333-b648-f4f20719e319","arxiv_id":"2501.14593","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":1,"one_line_summary":"A geometric-mean loss over softmax attention weights improves few-shot classification accuracy over arithmetic-mean losses.","lead":"This paper introduces a new loss function for few-shot image classification that uses the geometric mean, rather than the arithmetic mean, to combine similarities between a query image and the few labeled examples. It reports consistent accuracy gains over standard prototypical-network and neighborhood-component-analysis losses on three benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim is not yet isolated from the distance-metric choice: p=1 is selected via Fig. 1 on the proposed loss only, and no ablation shows the geometric-mean term, rather than L1, drives the gains over NCA/PN.","rationale":"The reader's weakest_assumption correctly identifies the missing ablation between the geometric-mean mechanism and the L1 metric choice. I agree that this is the most load-bearing gap, but I would sharpen it: it is not merely that the paper does not ablate the mechanism; it is that the only direct comparison of losses (Table 2) is made under a distance metric selected using the proposed loss alone, and one of the two theoretical motivations (the variance decomposition in Eq. 11) is derived for L2 while the experiments use L1. This makes the empirical claim 'geometric mean improves the loss' depend on an untested assumption that p=1 is equally optimal for all three losses. The mathematical derivation of Eq. (7) from Eq. (6) is correct, the upper-bound argument in Eq. (8) is correct, and the consistent gains across datasets are suggestive. However, because the distance-metric confound directly affects the headline comparison, the paper should be conditional on a clean ablation. I do not see a deeper internal inconsistency, and the theoretical core appears sound; the requested experiment could plausibly confirm the paper's claim, so a reject is not warranted. The verdict remains CONDITIONAL, hence UNCHANGED relative to the reader's assessment.","tokens_in":8225,"tokens_out":3885,"duration_ms":37300,"concrete_test":"Run a 2x2 ablation on miniImageNet 1-shot under the exact protocol of Sec. 3.1: {Ours, NCA} x {p=1, p=2}, keeping all other training details identical. If Ours with p=2 does not beat NCA with p=2 by at least the reported margin, or if NCA with p=1 matches Ours with p=1, then the headline improvement is attributable to the metric choice rather than geometric mean. Also report PN in the same grid and include the p-selection on validation classes rather than test classes.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central empirical claim is that Eq. (6) outperforms the PN loss (1) and NCA loss (2) in Table 2. But Table 2 is run entirely with dp=1 (L1), chosen from Fig. 1, which evaluates the distance metric only for the proposed loss. The baselines are never tuned for p; if their optimal p differs, the comparison is biased. More importantly, no experiment varies the loss-construction and the metric independently: e.g., Ours with dp=2 vs NCA/PN with dp=2 is not reported, nor is NCA with dp=1 vs Ours with dp=1 under identical embedding/epoch protocol except the geometric-mean/arithmetic-mean term. The theoretical support also has a metric mismatch: the within-class variance-decomposition in Eq. (11) is specific to Euclidean squared distance, while all headline experiments use p=1, where that decomposition does not hold and the stated medoid interpretation is different. Thus the claimed superiority of geometric mean over arithmetic mean is confounded with the switch to L1; the sparse-attention mechanism in Sec. 2.3.1 could in principle be tested with both p values.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a few-shot metric-learning loss that replaces the arithmetic mean of softmax-based attention weights in the NCA loss with a geometric mean (Eq. 6). The loss simplifies to an average distance plus a log-sum-exp term (Eq. 7). The authors provide theoretical analysis relating the loss to the NCA loss (upper bound and gradient weighting), to the PN loss (a within-class variance decomposition for squared L2 distance), and to multi-label classification. Experiments on miniImageNet, CIFAR-FS, and tieredImageNet with a ResNet12 backbone report consistent gains over the PN and NCA losses, along with comparisons to other few-shot methods.","tokens_in":8563,"tokens_out":5537,"duration_ms":53584,"significance":"The proposal is simple, parameter-free apart from the distance metric order p, and theoretically transparent; the derivations are easy to check and the computational overhead over the NCA loss is negligible. If the reported gains are robust and attributable to the geometric-mean aggregation, the loss would be a useful drop-in component for metric-based few-shot learning. The paper also correctly emphasizes the anti-sparsity behavior of the geometric mean. However, the central empirical claim is currently confounded with the choice of distance metric p=1, and part of the theoretical motivation is written for squared L2 distances while the headline experiments use L1 distances. The scientific value hinges on additional ablations that isolate the geometric-mean mechanism.","major_comments":[{"comment":"The central comparison to PN and NCA is run entirely with dp=1, a value selected from Fig. 1, which evaluates the distance metric only for the proposed loss. There is no analogous p-sweep for the PN and NCA baselines, and no experiment that varies the aggregation function (arithmetic vs. geometric mean) while holding the distance metric fixed. It is therefore possible that part of the reported gain (e.g., 65.51% vs. 62.68% on miniImageNet 1-shot) comes from imposing a suboptimal metric on the baselines rather than from the geometric-mean term. Please report Ours, PN, and NCA under both dp=1 and dp=2 under an otherwise identical training protocol, and preferably also include an ablation where the arithmetic-mean baseline is replaced by the geometric-mean term with the same p.","section":"Sec. 3.2, Table 2"},{"comment":"The variance-decomposition argument that the proposed loss additionally minimizes within-class variance is derived specifically for d(x,z)=||x-z||^2. The experimental section, however, adopts dp=1 for all headline results in Table 2, for which Eq. (11) does not hold. Thus the theoretical mechanism stated in Sec. 2.3.2 does not directly support the empirical configuration used in the main comparison. The paper should either supply a p=1 analogue of the variance/medoid analysis and connect it to the experiments, or run the headline comparison with p=2 to align theory and experiment.","section":"Sec. 2.3.2, Eq. (11)"},{"comment":"The claimed advantage of avoiding sparse attention weights is motivated analytically and by analogy to focal loss, but it is never empirically isolated. A direct test could compare the proposed loss against an arithmetic-mean NCA loss augmented with a simple anti-sparsity regularizer (e.g., adding -log a_i for in-class samples or an entropy term), under the same distance metric. Without such an ablation, the reader cannot tell whether the gains are due to the geometric-mean aggregation itself or to some other consequence of the loss form.","section":"Sec. 2.2 / Sec. 2.3.1"}],"minor_comments":[{"comment":"The inequality is attributed to the Cauchy-Schwarz inequality, but the direct justification is the AM-GM inequality (or Jensen's inequality on -log). The inequality itself is correct; the attribution should be fixed.","section":"Sec. 2.3.1, Eq. (8)"},{"comment":"Figure 1 shows performance versus p but does not report whether error bars or multiple runs are included. Please describe the protocol used to generate Fig. 1, including the number of trials and whether validation or test performance is plotted.","section":"Sec. 3.2, Fig. 1"},{"comment":"The text states that 'all the methods are equipped with dp=1' but does not explicitly say whether the PN and NCA baselines were re-trained under exactly the same leave-one-out mini-batch protocol, number of epochs, and learning-rate schedule as the proposed method. Please state this explicitly to confirm that the comparison is controlled.","section":"Sec. 3.2, Table 2"},{"comment":"The comparison methods [15-18] are cited from their respective papers, which may use different backbones, training protocols, or evaluation details; the underline criterion is also not fully defined. This is acceptable for context, but the text should acknowledge that Table 3 is not a controlled comparison.","section":"Sec. 3.3, Table 3"},{"comment":"There are several typographical errors: 'prose' should be 'propose' in Sec. 1, 'contraast' in Sec. 3.2, and reference [19] should use proper diacritics for 'João'. A careful proofread is recommended.","section":"Throughout"},{"comment":"The statement that p=1 'provides a medoid robust to outliers' is reasonable, but for p=1 the minimizer of the sum of distances is a geometric median, not necessarily a discrete medoid. The wording could be refined to avoid conflating the two notions.","section":"Sec. 2.4"}],"recommendation":"major_revision","confidential_remarks":"The paper is concise and appears to be in scope for a computer vision venue. The main technical idea is simple and potentially useful, but the missing ablations are load-bearing for the central claim: without isolating the geometric-mean aggregation from the distance-metric choice, the headline improvements over PN/NCA are not yet fully established. The theoretical/metric mismatch in Sec. 2.3.2 also needs to be addressed. These are fixable with additional experiments and reframing, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The geometric-mean loss in Eq. (6) is a simple and interesting modification of NCA, and the theory in Sec. 2.3 is mostly correct and illuminating. But the central empirical claim—that the geometric mean beats the arithmetic mean—is not cleanly isolated from the switch to L1 distance, which is chosen using only the proposed loss. Table 2 could therefore reflect metric preference rather than the loss form.\n\nWhat's good: the loss simplifies to mean in-class distance plus log-sum-exp over all distances, so it costs the same as NCA. The AM-GM upper bound is correct, and the gradient analysis in Eqs. (9)-(10) clearly shows why the loss up-weights far in-class samples. The multi-label interpretation and the medoid view are nice. The experimental framework follows standard leave-one-out mini-batch training and gives consistent 1-shot gains across all three benchmarks. Table 1 shows Ours beats BCE and ASL under L2, which is at least partial evidence that the geometric mean helps beyond the metric.\n\nThe main soft spot: no ablation isolates the loss aggregation from the distance metric. Fig. 1 selects p=1 by evaluating only the proposed loss; Table 2 then forces NCA and PN to use p=1 without checking whether that suits them. No experiment compares Ours vs NCA/PN under L2, or NCA under L1 vs Ours under L1 with everything else identical. Also, the variance decomposition in Eq. (11) holds only for L2, but the experiments use L1, so the theoretical support does not directly cover the reported setting. The claimed mechanism—avoiding sparse attention—is not directly tested. Minor: on tieredImageNet 5-shot the Ours-NCA gap is within confidence intervals.\n\nThis is a promising paper for the few-shot metric learning community. The loss is worth knowing even if the empirical claim needs tightening. A serious referee should ask for a proper ablation isolating the loss form from the distance metric, and ideally code. I'd send it to review, but with that request.","headline":"A genuinely neat loss with correct theory, but the headline empirical claim is confounded with the L1 distance choice; needs an ablation before the numbers are taken at face value.","tokens_in":8951,"tokens_out":2875,"would_cite":true,"duration_ms":26028,"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":"Replacing the arithmetic mean in a softmax few-shot loss with a geometric mean yields consistently better feature metrics.","keywords":["few-shot learning","metric learning","geometric mean","softmax attention","loss function","prototypical network","neighborhood component analysis","image classification"],"falsifier":"Train the proposed loss and the NCA loss under all four combinations of distance $p=1$ versus $p=2$ and arithmetic- versus geometric-mean aggregation on miniImageNet 1-shot. The attribution to the geometric mean is falsified if the geometric-mean loss does not beat NCA when both use $p=2$, or if NCA with $p=1$ matches the geometric-mean loss.","tokens_in":8006,"feed_emoji":"🎯","tokens_out":6095,"duration_ms":46990,"temperature":0.7,"pith_summary":"This paper proposes a loss for few-shot image classification that replaces the arithmetic mean over same-class softmax attention weights, the operation behind the NCA loss, with a geometric mean. Because a geometric mean collapses to zero if any attention weight does, the loss keeps every in-class sample active in the gradient even when it is far from the query. The paper provides three theoretical readings of this change: the loss is an upper bound of the NCA loss, its per-sample gradient weights are inversely proportional to attention rather than uniform, and it minimizes within-class variance in addition to center distance. On miniImageNet, CIFAR-FS, and tieredImageNet, the proposed loss beats the PN and NCA losses in all reported 1-shot and 5-shot settings, with the 1-shot gain on miniImageNet reaching 65.51 percent against 62.68 percent for NCA.","feed_headline":"Geometric mean loss beats few-shot baselines on all 12 settings","feed_subtitle":"Swapping arithmetic for geometric mean in softmax attention lifts 1-shot accuracy from 62.68 to 65.51 percent.","key_machinery":"The machinery is the geometric mean of softmax attention weights over samples of the query's class, $\\left[\\prod_{i:y_i=y_q}a_S(x_q,x_i)\\right]^{1/n_{y_q}}$. This is the operation that turns the NCA loss's arithmetic aggregation into the proposed loss. A geometric mean is zero whenever any factor is zero, so minimizing it pushes every in-class attention weight away from zero; the gradient analysis makes this explicit by weighting each sample-wise gradient term by $1/a_S(x_q,x_i)$, which assigns larger updates to far-away samples. The simplified form separates an average in-class distance from a log-sum-exp over all support samples, connecting the loss both to prototypical-net center learning and to NCA neighborhood weighting.","core_discovery":"The central claim is that aggregating softmax-based attention weights, $a_S(x_q,x_i)=\\exp(-d(x_q,x_i))/\\sum_j\\exp(-d(x_q,x_j))$, by geometric mean rather than arithmetic mean produces a feature metric that generalizes better in few-shot classification. The proposed loss is $\\ell_{\\mathrm{ours}}=-\\log\\left[\\prod_{i:y_i=y_q}a_S(x_q,x_i)\\right]^{1/n_{y_q}}$, which simplifies to $\\frac{1}{n_{y_q}}\\sum_{i:y_i=y_q}d(x_q,x_i)+\\log\\sum_j\\exp(-d(x_q,x_j))$. The paper argues that this drives a query toward the medoid of its class under the chosen distance, prevents sparse attention from silencing far-away in-class samples, and is equivalent to a multi-label softmax over the support set. In the reported comparisons, the losses use the $L_1$-based distance $d_1(x,z)=\\sum_{r}|x_r-z_r|$, which the paper selects after evaluating several $L_p$ variants.","pith_inferences":["The paper does not isolate the geometric mean from the L1-distance choice; a direct ablation of the two factors would settle which ingredient drives the reported gains.","The inverse-attention weighting in the gradient resembles an automatic focal-like reweighting, and a soft focal-loss variant might reproduce the same effect without a geometric mean.","The medoid-based interpretation could transfer to other metric-learning settings, such as retrieval or clustering, where a tolerant aggregate of pairwise similarities is desirable.","Because training uses a leave-one-out expectation over mini-batches, the loss's behavior in very small support sets might be sensitive to mini-batch composition; varying support size during training could test the robustness of the gains."],"forward_implications":["The geometric-mean loss can replace the NCA loss in an existing few-shot training loop at no extra computational cost, since both use distance terms plus a log-sum-exp.","Because it upper-bounds the NCA loss, the proposed loss keeps producing gradients in regimes where NCA is already saturated.","The medoid interpretation suggests the loss is tolerant of outlier support samples, especially when combined with the L1 distance.","On the three benchmarks, the loss alone is competitive with recent few-shot methods that add architectural components, despite changing only the training objective."],"supporting_citations":[{"why":"Supplies the prototypical-network loss that the proposed method is compared against and shown to exceed.","marker":"[5]"},{"why":"Provides the NCA loss whose arithmetic mean the paper replaces with a geometric mean; this is the central baseline and the source of the formulation.","marker":"[6]"},{"why":"Offers the focal-loss analogy for losses that remain effective after saturation, used to motivate the upper-bound property.","marker":"[10]"},{"why":"Supplies the mean-shift interpretation used to explain why inverse-attention gradient weighting helps metric learning.","marker":"[11]"},{"why":"Provides the medoid and discriminant-analysis notions used to characterize the L1 distance and within-class variance terms.","marker":"[12]"},{"why":"One of the multi-label classification losses compared in Table 1.","marker":"[13]"},{"why":"The asymmetric multi-label loss compared in Table 1.","marker":"[14]"}],"fun_headline_variants":["Geometric mean loss tops few-shot leaderboards","Swap arithmetic for geometric mean in loss, gain accuracy","Geometric mean in loss boosts few-shot performance","New loss: geometric mean beats softmax average","Geometric mean loss outperforms in few-shot tasks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper assumes that preventing any in-class attention weight from collapsing to zero is the mechanism behind the gains, but this mechanism is never ablated, and the reported improvement could partly come from the L1 distance chosen in Figure 1 rather than from the geometric mean itself.","fun_headline_variants_meta":{"raw":{"variants":["Geometric mean loss tops few-shot leaderboards","Swap arithmetic for geometric mean in loss, gain accuracy","Geometric mean in loss boosts few-shot performance","New loss: geometric mean beats softmax average","Geometric mean loss outperforms in few-shot tasks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000177,"raw_usage":{"total_tokens":1301,"prompt_tokens":961,"completion_tokens":340,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":577,"completion_tokens_details":{"reasoning_tokens":267}},"tokens_in":577,"tokens_out":340,"duration_ms":3991,"temperature":1.0,"reasoning_tokens":267,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T14:58:28.319338+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the proposed loss and the NCA loss under all four combinations of distance $p=1$ versus $p=2$ and arithmetic- versus geometric-mean aggregation on miniImageNet 1-shot. The attribution to the geometric mean is falsified if the geometric-mean loss does not beat NCA when both use $p=2$, or if NCA with $p=1$ matches the geometric-mean loss.","supporting_citations":[{"cited_title":"Generalizing from a few examples: A sur- vey on few-shot learning,","cited_arxiv_id":null,"evidence_quote":"Supplies the prototypical-network loss that the proposed method is compared against and shown to exceed."},{"cited_title":"Model-agnostic meta-learning for fast adaptation of deep networks,","cited_arxiv_id":null,"evidence_quote":"Provides the NCA loss whose arithmetic mean the paper replaces with a geometric mean; this is the central baseline and the source of the formulation."},{"cited_title":"On episodes, pro- totypical networks, and few-shot learning,","cited_arxiv_id":null,"evidence_quote":"Offers the focal-loss analogy for losses that remain effective after saturation, used to motivate the upper-bound property."},{"cited_title":"Matching net- works for one shot learning,","cited_arxiv_id":null,"evidence_quote":"Supplies the mean-shift interpretation used to explain why inverse-attention gradient weighting helps metric learning."},{"cited_title":"A discriminative feature learning approach for deep face recognition,","cited_arxiv_id":null,"evidence_quote":"Provides the medoid and discriminant-analysis notions used to characterize the L1 distance and within-class variance terms."},{"cited_title":"Neighbourhood components analy- sis,","cited_arxiv_id":null,"evidence_quote":"One of the multi-label classification losses compared in Table 1."},{"cited_title":"Focal loss for dense object detec- tion,","cited_arxiv_id":null,"evidence_quote":"The asymmetric multi-label loss compared in Table 1."}],"review_version":1}