{"id":"af91c7c5-7a3b-42ca-b490-2fb7c03a211d","arxiv_id":"2501.18619","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"FAAGC augments scarce training data by fitting and sampling a per-class geodesic arc in the pre-shape space, producing modest accuracy gains over prior feature-augmentation baselines.","lead":"This paper proposes a feature-level data augmentation method, FAAGC, that maps neural network features onto a sphere and creates synthetic training features by sampling along a per-class geodesic curve. It reports modest accuracy gains in few-shot image classification and a large speedup over the earlier FAGC method.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table I contradicts the 'consistently outperforms all baselines' claim: FAGC beats FAAGC on Caltech101@5 (88.72 vs 88.12) and on CIFAR-100@5 MLP (66.28 vs 66.23).","rationale":"The paper's central claim is that FAAGC is uniformly better than prior feature-level augmentation methods under data scarcity, especially FAGC. My check of the reported tables found direct counterexamples within the manuscript. The reader's weakest assumption about geodesic fit is scientifically important, but it is not the most load-bearing issue because the empirical claim fails on the paper's own numbers even before any geometric assumption is tested. Table I gives FAGC 88.72±0.03 vs FAAGC 88.12±0.01 on Caltech101@5; Table II gives FAGC 66.28 vs FAAGC 66.23 for MLP on CIFAR-100@5. Since no code or per-seed values are available, the small standard deviations cannot be independently verified. This is an internal inconsistency rather than a disagreement with outside consensus, so it should be fixed by revising the claim and releasing evidence; it does not by itself falsify the augmentation idea. Hence I keep the reader's CONDITIONAL verdict, which already asks for corrections, and mark no verdict change.","tokens_in":15353,"tokens_out":7909,"duration_ms":80820,"concrete_test":"Obtain the authors' code and per-seed logs, then rerun the Caltech101@5 experiment in Table I and the CIFAR-100@5 MLP experiment in Table II under the exact stated protocol and six seeds, computing paired per-seed differences between FAAGC and FAGC. If FAGC wins either comparison beyond seed noise, the paper must qualify or remove the 'consistently outperforms all baselines' assertion; if FAAGC wins both, the contradiction is a reporting error.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim in Section IV-B that 'FAAGC consistently outperforms all baseline methods under data-limited conditions across all benchmark datasets' is not supported by the paper's own tables. In Table I, on Caltech101@5, FAGC achieves 88.72±0.03 while FAAGC achieves 88.12±0.01, a 0.60-point deficit that exceeds the reported standard deviations. In Table II, on CIFAR-100@5 with an MLP classifier, FAGC obtains 66.28 vs FAAGC 66.23. These are exactly the comparisons the paper uses to argue FAAGC improves on the prior geodesic method FAGC. The claim of universal superiority therefore fails on the reported evidence, independent of any geometric assumption about the pre-shape space. This does not show the method is useless; it means the headline assertion is overstated and needs qualification. The reader's concern about untested geodesic low-dimensionality is real but secondary: even if class features are well approximated by a geodesic arc, the reported numbers already contradict 'consistently outperforms all baselines.' No code or per-seed results are provided, so the small standard deviations cannot be checked.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FAAGC, a feature-level data augmentation method that projects deep image features into a pre-shape space by duplicating each feature coordinate, centering, and normalizing onto a unit sphere. For each class, it fits an adaptive geodesic arc by optimizing two endpoint vectors and a set of sampling parameters with a loss that combines a sample-similarity term and a distribution-divergence term, then augments by sampling uniformly along the fitted arc. Experiments on CIFAR-10, CIFAR-100, CUB-200, Fashion-MNIST, and Caltech101 under 5-samples-per-class conditions compare FAAGC with several feature augmentation baselines and with the prior FAGC method, reporting accuracy gains in several settings and a large runtime reduction.","tokens_in":15692,"tokens_out":6800,"duration_ms":67702,"significance":"If the results are taken at face value, the main contributions are a gradient-based geodesic fitting procedure that is roughly an order of magnitude faster than FAGC (Table V) and consistent gains over no augmentation and over most baselines in low-data settings (Tables I and IV). The method is simple, modular, and the appendix proof of equivalence between the geodesic and interpolation forms is mathematically sound. However, the paper's headline claim that FAAGC 'consistently outperforms all baseline methods' is not supported by its own tables, and the algorithm description contains an internal inconsistency that must be resolved before the results can be reproduced. No code or per-seed results are provided, so the small reported differences cannot be independently checked.","major_comments":[{"comment":"The claim that 'FAAGC consistently outperforms all baseline methods under data-limited conditions across all benchmark datasets' is directly contradicted by the reported numbers. In Table I, on Caltech101@5, FAGC achieves 88.72±0.03 while FAAGC achieves 88.12±0.01, a deficit of 0.60 points that exceeds the reported standard deviations. In Table II, on CIFAR-100@5 with an MLP classifier, FAGC obtains 66.28 while FAAGC obtains 66.23. In Table III, on CIFAR-100@5 with Swin-Transformer, Manifold-Mixup obtains 73.34 while FAAGC obtains 73.26. These are not incidental comparisons: FAGC is the prior geodesic method that FAAGC claims to improve, and Manifold-Mixup is a central baseline. The central claim must be revised to state precisely in which settings FAAGC wins, and the text should discuss the cases where it does not.","section":"Section IV-B, Tables I-III"},{"comment":"There is an internal inconsistency about whether the sampling parameter t is resampled or learned. The text states that 'the sampling parameters are treated as a set of learnable parameters of size m' and that optimization updates t, but Algorithm 1, line 2, says 'Sample t ∼ N(0, I)' inside the while loop, and line 4 applies a sigmoid. If t is re-sampled at every iteration, there is no persistent parameter for Adam to update at line 10, and the divergence loss L_Diverg cannot align the learned t with the final augmentation distribution z ∼ U(0,1). If t is initialized once and then updated, line 2 should say 'Initialize t'. The authors must specify the exact training procedure, since the method's reproducibility hinges on this distinction.","section":"Section III-B and Algorithm 1"},{"comment":"The geometric premise of the method is that each class's pre-shape features are well approximated by a single geodesic arc, and that sampling points on that arc produces semantically valid class members. This premise is load-bearing but is never tested. The fitting loss L_Sim is minimized, but no diagnostic is reported for the residual geodesic distance after fitting, nor is there any comparison with alternative low-dimensional models such as principal geodesic analysis or multiple arcs. Without such evidence, the uniform sampling along a single fitted arc may generate off-manifold features. I recommend adding a quantitative test that reports, for each dataset and class, the fit quality of the learned geodesic relative to the spread of the training features.","section":"Section III-A and III-B"},{"comment":"The hyperparameters β, η_p, and η_t are selected by grid search on the reduced CIFAR-100 training set and then applied uniformly to all datasets and backbones. Since CIFAR-100@5 is also the dataset on which the largest accuracy improvement is reported (1.46 points in Table I), the evaluation does not fully separate hyperparameter choice from method effectiveness. Please add an analysis showing sensitivity across datasets when the hyperparameters are tuned per dataset, or justify that the chosen values are not the source of the reported gains.","section":"Section IV-A and IV-B"}],"minor_comments":[{"comment":"The no-augmentation baselines are reported with ±0.00 standard deviation, despite the statement that 'all later experiments are conducted 6 times with different random seeds.' Please clarify whether these baselines are deterministic or were run once; if they were run multiple times, the zero standard deviation is suspicious.","section":"Table I"},{"comment":"The training times for FAAGC are 39.39 s for 5 samples per class and 39.34 s for 10 samples per class, which are almost identical. This is surprising because the fitting workload should scale with m. Please explain whether the timing includes a fixed overhead or whether the measurements are noisy, and report standard deviations for the runtime.","section":"Table V"},{"comment":"The up-dimension procedure duplicates each feature value to form landmark pairs, i.e., v'_i = [(v_i[1], v_i[1]), ..., (v_i[d], v_i[d])]. After centering by the mean of each coordinate, the x and y coordinates of every landmark are identical, so all landmarks lie on the line y=x. This makes the configuration degenerate as a 2D landmark set. Please clarify whether this degeneracy is intentional and how it affects the interpretation of the pre-shape space projection.","section":"Section III-A"},{"comment":"There are several typos and grammatical errors, such as 'a object' in Section II-B and 'featires' in Section III-C. A careful proofread is needed.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"This is a borderline paper. The runtime improvement over FAGC is concrete and the general idea is interesting, but the central accuracy claim is overstated and Algorithm 1 is internally inconsistent. The material can likely be fixed with a revised exposition, qualification of the claims, and additional validation of the geodesic assumption. I recommend asking for code and per-seed results, since the reported differences are often smaller than typical seed variability and there is no way to verify them from the paper alone."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this is FAGC plus gradient-based curve fitting, learnable sampling parameters, and a Wasserstein divergence loss. The runtime improvement is the most convincing part: Table V shows a drop from 351s to 39s on CIFAR-10@5 and 711s to 39s on CIFAR-10@10. That is a real engineering contribution, and the matrix-form loss computation is a sensible way to get there.\n\nWhat is genuinely new is modest but legitimate. The learnable t with a divergence loss addresses a real correspondence problem during curve fitting—the paper explains that clearly with Figure 2—and the gradient-based fitting is a natural replacement for FAGC's iterative geodesic distance search. The appendix proof that the geodesic and interpolation forms are equivalent is correct and useful.\n\nThe soft spots are substantial. First, the paper states in Section IV-B that FAAGC \"consistently outperforms all baseline methods,\" but Table I shows FAGC beating FAAGC on Caltech101@5 (88.72 vs 88.12), and Table II shows FAGC beating FAAGC on CIFAR-100@5 with an MLP (66.28 vs 66.23). These are exactly the comparisons against the prior geodesic method, so the headline claim fails on the reported evidence. Second, Algorithm 1 conflicts with the text: the text says t is a set of learnable parameters of size m, but the algorithm samples t from N(0,I) in every iteration before the sigmoid. That ambiguity sits at the core of the fitting procedure and needs to be resolved. Third, there is no code, no per-seed results, and several tables (II, III, IV) lack error bars entirely, so the small standard deviations in Table I cannot be checked. Fourth, the load-bearing geometric assumption—that class features lie near a single geodesic arc in pre-shape space—is never tested. The stress-test note is right that this is secondary to the overclaim, but a diagnostic would be valuable.\n\nWho is this for? People doing feature-level augmentation in low-data settings, especially those who found FAGC too slow. The citation pattern is honest: the method leans on the authors' own FAGC for the up-dimension and loss weighting, and that is disclosed. It deserves a serious referee because the speedup and the learnable-t idea are worth checking, but the current version overclaims and has a central algorithmic ambiguity.\n\nSend it to peer review with the expectation of major revision: fix the algorithm description, qualify the consistency claim, and provide code or at least per-seed numbers. I would not cite it in its current form.","headline":"Useful incremental follow-up on FAGC with a real speedup, but the headline claim of consistent superiority is contradicted by the paper's own tables; referee it, expecting major revision.","tokens_in":16165,"tokens_out":1648,"would_cite":false,"duration_ms":16109,"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":"FAAGC claims that sampling synthetic features along a per-class geodesic arc in pre-shape space improves few-shot classification, reporting gains on five benchmarks and a roughly 9x speedup over the earlier FAGC.","keywords":["Feature augmentation","Shape space theory","Pre-shape space","Geodesic curve","Data scarcity","Few-shot learning","Image classification","Gradient descent"],"falsifier":"Fit the FAAGC geodesic for each class on a few-shot training set and measure the mean geodesic distance from every training feature to the fitted arc, or train a binary discriminator to separate real features from FAAGC-sampled features. If the residuals are large or the discriminator easily separates the two, for instance with a simple linear probe achieving above-chance accuracy, the arc does not represent the class and the augmented points are off-manifold, which would invalidate the claimed mechanism.","tokens_in":15185,"feed_emoji":"📈","tokens_out":7631,"duration_ms":61183,"temperature":0.7,"pith_summary":"FAAGC is a feature-level data augmentation method for deep classifiers trained with very few examples per class. The paper claims that projecting each class's features into the pre-shape space, the unit sphere obtained after centering and normalizing, and fitting an adaptive geodesic arc to them yields synthetic samples that stay closer to the true feature distribution than linear interpolation or noise-based augmentation. Across five image benchmarks reduced to 3–20 samples per class, the method reports consistent accuracy gains, with the largest improvement on CIFAR-100@5 (67.87% vs 66.41% without augmentation). It also claims a large speedup over the earlier FAGC approach, cutting geodesic-curve fitting time from about 352 seconds to 39 seconds on CIFAR-10@5. If correct, the contribution is a generic, classifier-agnostic plug-in that makes scarce-data training more reliable without hand-designed image transforms.","feed_headline":"Geodesic-curve sampling boosts accuracy when data is scarce","feed_subtitle":"Fitting one arc per class beats linear interpolation in few-shot tests and runs 9x faster than prior method.","key_machinery":"The machinery is the pre-shape space construction and the adaptive geodesic arc. A feature vector $v\\in\\mathbb{R}^d$ is expanded to paired coordinates $(v[1],v[1]),\\dots,(v[d],v[d])$, centered and normalized to a unit-sphere vector $\\tau\\in S^{2d-3}_*$; by shape space theory, all objects of one shape lie on a great circle, so a class is represented by an arc $\\Gamma(\\tau_{\\text{start}},\\tau_{\\text{end}})(z)=\\frac{\\sin[(1-z)\\theta]}{\\sin\\theta}\\tau_{\\text{start}}+\\frac{\\sin(z\\theta)}{\\sin\\theta}\\tau_{\\text{end}}$, with $\\theta$ the geodesic distance between endpoints and $z\\in[0,1]$ the normalized arc position. The endpoints are learned per class: the loss $L_{\\text{train}}=L_{\\text{Sim}}+\\beta\\,L_{\\text{Diverg}}$ aligns sampled points with real features by cosine similarity and forces the learnable positions $t$ to match $U(0,1)$ by sorted Wasserstein distance, while the endpoints are kept on the sphere by optimizing the unnormalized Euclidean vectors and projecting. This is what carries the argument: the fitted arc is the generative model, and sampling along it is the augmentation.","core_discovery":"The central claim is that, for each class, the deep features of a small training set lie close to a single geodesic arc in the pre-shape space, and that sampling new features along that arc creates useful training data. The paper implements this by duplicating each feature coordinate to form 2D landmarks, subtracting the mean, and normalizing onto the unit sphere $S^{2d-3}_*$; it then learns two endpoint vectors per class by gradient descent on a loss that combines a cosine-similarity term between sampled and real features with a Wasserstein term that keeps learnable sampling parameters close to the uniform distribution used at generation time. After fitting, augmented features are sampled from $\\Gamma(\\tau_{\\text{start}},\\tau_{\\text{end}})(z)$ with $z\\sim U(0,1)$. On five benchmarks at 5 samples per class, FAAGC reports the best accuracy among all compared feature-level augmentation methods, including a 1.46-point gain over the no-augmentation baseline on CIFAR-100@5, and it outperforms FAGC while reducing fitting time from 351.83s to 39.39s on CIFAR-10@5.","pith_inferences":["(Editorial inference) The single-arc-per-class model should be testable directly: fitting FAAGC on a class with two well-separated visual modes, such as a digit written with open or closed top, would likely produce a high residual loss, predicting that a mixture of two geodesic arcs would outperform one.","(Editorial inference) The projection duplicates every coordinate to synthesize 2D landmarks; if the up-dimension step distorts the geometry, alternative embeddings, such as pairing consecutive coordinates or using a spherical map of the original vector, may change augmentation quality, and a controlled comparison would isolate the projection's role.","(Editorial inference) The reported speedup makes online augmentation during classifier training feasible, a setting the paper leaves unexplored, since the current pipeline augments features offline before classifier training.","(Editorial inference) The method's success suggests that pre-trained deep features of a class, in the low-data regime, concentrate near low-dimensional great-circle arcs; if so, principal geodesic analysis or multi-geodesic mixtures could yield further gains without new data."],"forward_implications":["At 5 samples per class, FAAGC reports the highest accuracy among compared methods on all five datasets; the largest gain over no augmentation is +1.46 points on CIFAR-100@5.","Gains are largest when data is scarcest: at 3 samples per class CIFAR-100 improves from 57.83% to 59.39%, while at 20 samples per class the gain shrinks to 0.26 points.","FAAGC improves accuracy for k-NN, SVM, and MLP classifiers, and across ViT, Swin, ResNet, and EfficientNet backbones, indicating it is not tied to one architecture.","The method stacks with image-level augmentation: combining FAAGC with RandAugment or AugMix on CIFAR-100@5 gives additional gains over either alone.","Compared with FAGC, FAAGC reduces fitting time from 351.83s to 39.39s on CIFAR-10@5 while improving accuracy from 85.20% to 85.84%."],"supporting_citations":[{"why":"Prior FAGC method that first built geodesic curves in pre-shape space for feature augmentation; FAAGC is presented as its faster, more accurate successor and the main comparison baseline.","marker":"[10]"},{"why":"Introduced the shape space theory, establishing that same-shape objects lie on great circles in pre-shape space, the geometric premise of the method.","marker":"[8]"},{"why":"Earlier work applying shape-space geodesics to object recognition with limited data; supplies the geodesic sampling formula used for augmentation.","marker":"[9]"},{"why":"Book-length treatment of shape and shape theory giving the geodesic distance metric and pre-shape space formalism used in the loss.","marker":"[23]"},{"why":"Seminal feature-level augmentation work that motivated synthetic features in feature space rather than input space; contextual baseline.","marker":"[5]"},{"why":"Manifold Mixup baseline representing linear interpolation of hidden features, which FAAGC claims to outperform.","marker":"[6]"},{"why":"Feature-level SMOTE baseline for feature-space augmentation, used as a comparison in experiments.","marker":"[7]"},{"why":"Defines the Wasserstein distance used in the divergence loss that aligns learnable sampling parameters with the uniform augmentation distribution.","marker":"[27]"}],"fun_headline_variants":["Geodesic arcs boost few-shot accuracy via feature sampling","Sampling along geodesic curves improves scarce-data classification","Adaptive geodesic curves improve classification with sparse data","Few-shot accuracy via feature sampling on class geodesics"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that each class's deep features, after duplicating coordinates, centering, and normalizing onto the unit sphere, lie close to a single geodesic arc, and that any point sampled on that arc still looks like a valid member of the class; the paper offers no geometric or statistical test of this assumption.","fun_headline_variants_meta":{"raw":{"variants":["Geodesic arcs boost few-shot accuracy via feature sampling","Sampling along geodesic curves improves scarce-data classification","Adaptive geodesic curves improve classification with sparse data","Few-shot accuracy via feature sampling on class geodesics"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000811,"raw_usage":{"total_tokens":3539,"prompt_tokens":909,"completion_tokens":2630,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":525,"completion_tokens_details":{"reasoning_tokens":2566}},"tokens_in":525,"tokens_out":2630,"duration_ms":17666,"temperature":1.0,"reasoning_tokens":2566,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T14:31:09.245027+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fit the FAAGC geodesic for each class on a few-shot training set and measure the mean geodesic distance from every training feature to the fitted arc, or train a binary discriminator to separate real features from FAAGC-sampled features. If the residuals are large or the discriminator easily separates the two, for instance with a simple linear probe achieving above-chance accuracy, the arc does not represent the class and the augmented points are off-manifold, which would invalidate the claimed mechanism.","supporting_citations":[{"cited_title":"FAGC:Feature Augmentation on Geodesic Curve in the Pre-Shape Space","cited_arxiv_id":"2312.03325","evidence_quote":"Prior FAGC method that first built geodesic curves in pre-shape space for feature augmentation; FAAGC is presented as its faster, more accurate successor and the main comparison baseline."},{"cited_title":"Shape manifolds, procrustean metrics, and complex projective spaces,","cited_arxiv_id":null,"evidence_quote":"Introduced the shape space theory, establishing that same-shape objects lie on great circles in pre-shape space, the geometric premise of the method."},{"cited_title":"Recognition of multiple configurations of objects with limited data,","cited_arxiv_id":null,"evidence_quote":"Earlier work applying shape-space geodesics to object recognition with limited data; supplies the geodesic sampling formula used for augmentation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Book-length treatment of shape and shape theory giving the geodesic distance metric and pre-shape space formalism used in the loss."},{"cited_title":"Feature-level smote: Augmenting fault samples in learnable feature space for imbal- anced fault diagnosis of gas turbines,","cited_arxiv_id":null,"evidence_quote":"Feature-level SMOTE baseline for feature-space augmentation, used as a comparison in experiments."}],"review_version":1}