{"id":"88dce118-3fac-468e-a1dd-579073d17a0b","arxiv_id":"2506.20850","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"COVER reformulates contrastive learning as vector regression of geometric displacement fields and reports state-of-the-art self-supervised pretraining results across 2D and 3D medical imaging tasks.","lead":"COVER trains medical image models to predict how pixels move between two transformed views of the same image, replacing the binary pull/push of standard contrastive learning. The authors report consistent gains over 17 baselines on 8 downstream medical tasks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The regression target v is a 2-D mean of the N×N distance map, so L_vec cannot constrain individual feature distances or dispersion; one-hot and broad distance maps can share identical loss, undermining the quantitative-dispersion claim.","rationale":"The paper's strongest claim is that vector CL models pixel-wise feature distances to quantify dispersion. For this to hold, minimizing L_vec must constrain the N×N distance map. It does not: the predicted vector is a 2-D weighted mean (or, in 3D, a 3-D weighted mean), so the regression is massively underdetermined. This is more basic than the representability issue the reader raised: even granting Eq. (21), the loss can be zero for many different distance maps, so individual distances are not 'modeled'. The supplement's bound also relies on α_min > 0, which the loss does not enforce; the degenerate one-hot solution is actually encouraged when the target lies on a template offset. I therefore partially agree with the reader's weakest assumption: the load-bearing issue is underdetermination rather than representability of the affine DVF. The empirical study is broad and reproducible, with public datasets, a code link, ablations, cross-architecture tests, and a test-retest reliability check. Those results are credible but do not rescue the theoretical framing. My recommendation is to keep the CONDITIONAL verdict: the method may work, but the title claim and the Sec. 3.1/Sec. 3.3 discussion should be reworded or augmented with a constraint or regularizer on the distance-map distribution before the vector-CL paradigm is presented as quantifying dispersion.","tokens_in":27894,"tokens_out":13419,"duration_ms":168872,"concrete_test":"Analytical check: for N=7, template offsets include (0,0), (1,0), and (2,0). Choose two softmax maps, D1 with weight 1 on (1,0) and D2 with weight 0.5 on (0,0) and 0.5 on (2,0). Both yield v'=(1,0), so L_vec is identical for ground truth v=(1,0), yet the spatial entropy/dispersion differs maximally. This demonstrates that Eq. (20) does not determine the distance map. Empirical follow-up: load a trained COVER, record the softmax maps D for a minibatch, and compute their variance/entropy; compare against DenseCL's similarity distributions. If the learned maps are not systematically broad, the pretraining objective is not the source of any dispersion control, and the central claim should be revised to 'soft-argmax alignment pretraining' without the dispersion-quantification interpretation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Eq. (7), VEU outputs v'_i = Σ_j softmax_j(f)·V_j, a spatial weighted mean of the template offsets. For an N×N=49-pixel receptive field, L_vec = |v'−v| imposes at most two scalar constraints, so there is a large family of softmax/distance maps with exactly the same predicted vector. The loss therefore cannot quantify dispersion: a one-hot distance map (all mass on a single neighbor, maximal over-dispersion) and a broad map with the same mean vector are indistinguishable to L_vec. This is not a corner case; for any v in the template hull, the preimage under V is high-dimensional. The theoretical argument in Suppl. A.2 assumes the softmax weights α_j approach given coefficients and derives δ_VCL ≤ τ log(1/α_min), but nothing in L_vec enforces α_j ≥ α_min > 0 or any spread of the weights. In fact, whenever the ground-truth DVF equals a template offset, L_vec is minimized by collapsing the attention to that single offset—exactly the binary over-dispersion the paper says it avoids. Thus the central assertion that COVER 'quantifies dispersion degrees' and preserves pixel-wise correlations is not supported by the objective; the implemented regression constrains only the centroid of the distance map. The broad empirical gains may still be real and useful, but they cannot be attributed to the claimed distance/dispersion modeling without an additional constraint on the distribution of D.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a new self-supervised pretraining paradigm for pixel-wise medical representation learning, called vector contrastive learning (VCL), and instantiates it in the COVER framework. COVER generates two views of an image via appearance and spatial affine transformations, uses the known displacement vector field (DVF) as a free regression target, and predicts the DVF from dense feature distances through a vector embedding unit (VEU), a multi-vector integration (MVI), and a vector pyramid aggregation (VPA) architecture. The central theoretical claim is that regressing these displacement vectors is equivalent to modeling pixel-wise feature distances and therefore quantifies feature dispersion, avoiding the alleged over-dispersion of binary contrastive learning. The empirical section reports pretraining on 2D chest X-ray and 3D brain MRI, fine-tuning on 8 tasks spanning 4 modalities, and comparisons with 17 baselines, together with ablations, reliability analysis, and cross-architecture compatibility experiments.","tokens_in":28230,"tokens_out":3916,"duration_ms":48726,"significance":"If the central equivalence claim were correct, COVER would be a notable contribution: it would offer a principled, annotation-free way to turn distance modeling into vector regression, with a concrete mechanism for preserving pixel-wise feature correlations and a tighter generalization bound. The empirical protocol is genuinely broad and is a strength of the paper: 8 downstream tasks, both 2D and 3D settings, 4 imaging modalities, 17 baselines, component and hyperparameter ablations, and a public code link. The reported gains are consistent across tasks, and the method is simple enough to reproduce. However, the theoretical justification in Sec. 3.1.2 and in Supplementary A.2 is not valid as stated: the loss constrains only a low-dimensional centroid of the distance map, not the per-pair distances or their dispersion. The paper itself concedes (Sec. 3.1.2) that 'vector regression is not strictly equivalent to conventional distance modeling,' which is in tension with the abstract and Sec. 3.1's stronger claims. The empirical value of COVER may still stand, but the paper currently over-interprets its mechanism, and the novel theoretical framing needs substantial rework.","major_comments":[{"comment":"The claimed equivalence chain is not a valid derivation. Equation (2) asks every per-pair distance error |α_i - d'_i| to vanish; Eq. (3) asks the weighted norm of those errors to vanish. If the vectors V_i are not linearly independent (they are not in the VEU, because N×N > 2 in 2D), convergence of the weighted sum in Eq. (4) does not imply convergence of the individual terms in Eq. (3), so the double arrow between Eqs. (2) and (3) is false. Equation (4) itself only gives convergence of a weighted average, which is the same as Eq. (5). Thus the chain establishes only that the regressed vector approaches the ground truth vector v, not that feature distances approach the hypothesized ground truth distances α_i. The paper's own caveat, 'vector regression is not strictly equivalent to conventional distance modeling,' correctly identifies this gap, but that caveat undercuts the paper's central claim as stated in Sec. 3.1 and the abstract.","section":"Sec. 3.1.2, Eqs. (2)-(5)"},{"comment":"The VEU output is a convex combination of the fixed template vectors: v'_i = Σ_j softmax_j(f)·V_j with softmax weights summing to 1. Consequently L_vec in Eq. (6) imposes at most two scalar constraints (for 2D) on the N×N distance map. For any ground-truth vector v in the convex hull of the template, there is a high-dimensional family of softmax/distance maps with exactly the same predicted vector, including one-hot maps (all mass on a single neighbor) and broad maps (mass spread over many neighbors). This means L_vec cannot quantify dispersion: a one-hot distance map and a broad map with the same centroid are indistinguishable to the loss. The claim in Sec. 3.1.1 that vector regression 'quantifies their dispersion degrees' is therefore not supported by the implemented objective. The authors need to either add an explicit concentration/spread penalty on the distance map or substantially revise the dispersion-quantification claims.","section":"Sec. 3.3, Eq. (7)"},{"comment":"The theory assumes that the ground-truth displacement vector v can be written as v = Σ_j α_j V_j with α_j ≥ 0 and Σ_j α_j = 1, and then derives δ_VCL ≤ τ log(1/α_min) from α_j ∈ [α_min, 1]. Nothing in L_vec enforces α_j ≥ α_min > 0, nor does anything prevent the model from collapsing the softmax weights onto a single template vector. In the common situation where the ground-truth DVF equals a template offset (e.g., pure translation by one pixel, or a scaled version of a template vector), the loss is minimized by putting almost all softmax mass on that single offset, which reproduces exactly the binary over-dispersion the paper says it avoids. The bound in Eq. (24) and the tighter generalization bound in Eq. (25) therefore do not follow from the objective. If the authors want to keep this theory, they must show that the optimization dynamics or an explicit regularizer keep the weights bounded away from the simplex vertices.","section":"Supplementary A.2, Eqs. (21)-(24)"},{"comment":"The empirical sections attribute the observed gains to 'distance modeling' and 'quantitative dispersion' (e.g., the component ablation and the t-SNE analysis). Because L_vec only constrains the centroid of the distance map, these attributions are not established by the current experiments. The reported improvements may be real and could result from the affine-prediction pretraining task, the multi-scale architecture, or the consistency loss, rather than from dispersion modeling. The authors should add a diagnostic that directly measures the dispersion of the predicted distance maps (e.g., entropy of the softmax distribution over the receptive field) under L_vec and under the compared baselines, or rephrase the claims to describe the method as centroid-based vector regression with multiscale fusion. Without such evidence, the mechanism-related conclusions in Secs. 4.2.1 and 4.3.6 are not supported by the objective.","section":"Sec. 4.2.1 and Sec. 4.3.6"}],"minor_comments":[{"comment":"The notation is garbled: 'PI i=0' appears to be an incomplete sum, and the limits and indexing are unclear (i = 1..I or i = 0..I). Please rewrite the chain with standard sum notation and define V_i and d'_i explicitly before the equations.","section":"Sec. 3.1.2, Eqs. (2)-(5)"},{"comment":"The pretraining dataset is called 'ChestX-ray6' here but 'ChestX-ray8' in the datasets table (Table 2) and elsewhere. This inconsistency should be fixed.","section":"Sec. 4.1.1"},{"comment":"The fusion operation J is described qualitatively in the main text and formally only in Supplementary B.4. The main text should at least define J when it is first used, since Eq. (8) is the core architecture equation.","section":"Sec. 3.4, Eq. (8)"},{"comment":"The column 'A VG' is not defined in the caption or text; if it is the average rank, say so explicitly. Also, the numbers in the COVER row (e.g., 63.6 without a '+' sign for FeTA21) should be formatted consistently.","section":"Table 1"},{"comment":"The text references 'Fig.8 in the manuscript' when referring to the multi-scale vector visualization, but the main manuscript's Fig. 8 is the vector template diagram. The cross-references between main text and supplementary figures need to be corrected.","section":"Supplementary D.5"},{"comment":"There are several grammatical and typographical errors, e.g., 'putted', 'the the', and '2×10 5 iterations'. A careful proofreading pass is needed.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper has a solid empirical backbone and the method could be a useful practical pretraining recipe. The core problem is that the theoretical framing in Sec. 3.1 and Supplementary A.2 overclaims what the loss actually does. I would like the revision to add either an explicit dispersion penalty that makes the stated mechanism true, or to remove the equivalence/quantification claims and reposition the paper as a centroid-based vector regression pretraining method. In the latter case the abstract and Sec. 3.1 should be rewritten accordingly, and the generalization-bound section should be removed or substantially weakened. The paper cites its own previous work (GVSL, GEMINI) heavily but that is not in itself problematic; however, the novelty claims should be re-examined against those same references after the revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a useful empirical paper with a central theoretical claim that does not survive contact with Eq. (7). The vector regression loss constrains the centroid of the distance map, not the dispersion, and the paper's own Section 3.1.2 concedes the equivalence is not strict. I don't think that sinks the work—the recipe seems to work—but the authors should not be allowed to publish the 'quantifies dispersion degrees' statement as is.\n\nWhat is genuinely new: framing contrastive learning as vector regression with a self-generated displacement vector field, plus the VEU, MVI, and VPA components. It builds on their own GVSL and GEMINI work, but the multi-scale pyramid and mixture-of-vectors are not trivial restatements. The empirical protocol is broad: 8 downstream tasks, 2D and 3D, 4 modalities, 17 baselines, all public datasets, and a code link. The ablations, cross-architecture compatibility test, reliability check, and FLOPs analysis are real evidence. The gains over GVSL and GEMINI are consistent enough to take seriously.\n\nThe soft spot is load-bearing. VEU outputs v' = softmax(f_a f_b^T) V, a weighted mean of template offsets. L_vec = |v' - v| imposes two scalar constraints per pixel on a 7x7 map of 49 similarity weights. One-hot and flat distance maps with the same centroid have identical loss. Nothing in the objective enforces alpha_min > 0 or any spread of attention weights. The supplement's bound assumes exactly that, but the loss does not require it; in fact, when the ground-truth DVF equals a template offset, collapsing attention to that single offset minimizes L_vec perfectly—the opposite of avoiding over-dispersion. The paper itself already admits that aggregated error does not imply per-pair convergence, but then proceeds as if dispersion is quantified anyway. So the theory is not just mildly loose; it does not support the central name and motivation. Also, the main comparison table has no error bars, and the two new hyperparameters N and J are tuned on the SCR linear evaluation target, so the headline numbers are somewhat cherry-picked. Those are fixable with independent validation.\n\nWho this is for: people building medical pixel-wise self-supervised pipelines will find this valuable as a recipe and a benchmark. It deserves a serious referee; an editor should send it out. I would ask the authors to either re-derive the distance-modeling connection with an explicit constraint on D, or reframe as an empirical correspondence-regression method and stop claiming dispersion quantification. Either path can yield a good paper; the current version is not there.","headline":"A broad, reproducible pixel-wise pretraining recipe whose empirical gains are plausible but whose motivating 'quantitative dispersion' story is not supported by the objective; worth reviewing, but the theory needs a major revision or removal.","tokens_in":28726,"tokens_out":2098,"would_cite":true,"duration_ms":25177,"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":"Contrastive learning for pixel-wise medical pretraining can be reformulated as vector regression, and the resulting COVER framework outperforms 17 baselines across 8 downstream tasks.","keywords":["vector contrastive learning","pixel-wise self-supervised pretraining","displacement vector field","over-dispersion","distance modeling","medical image segmentation","medical vision foundation models","contrastive learning"],"falsifier":"Observe the correlation between learned feature distances and true spatial displacements on transformations outside the template's range, such as affine shifts larger than the template size or elastic deformations. If the Pearson correlation stays high, the distance-modeling equivalence generalizes; if it drops while the vector regression loss remains low, the model is fitting vectors without actually modeling distances, falsifying the central claim.","tokens_in":27670,"feed_emoji":"🩻","tokens_out":7692,"duration_ms":79757,"temperature":0.7,"pith_summary":"This paper takes on a known failure mode of contrastive self-supervised learning when it is applied pixel by pixel to medical images: binary contrastive objectives push features apart without controlling how far, so nearby pixels that belong to the same anatomical structure end up over-dispersed and their correlations break. The proposed remedy is to reformulate contrastive learning as a vector regression problem: instead of maximizing and minimizing feature distances directly, the model predicts a dense displacement vector field between two randomly transformed views of the same image, with the known spatial transformation supplying free ground-truth vectors. The paper argues that regressing these vectors implicitly models feature distances, quantifying dispersion while preserving local correlations. It implements this idea in the COVER framework and reports that pretraining in this way beats 17 comparison methods across 8 downstream medical tasks in 2D and 3D. If the claim holds, pixel-wise self-supervised pretraining no longer needs binary contrast at all and can be driven entirely by geometric self-supervision.","feed_headline":"Pixel-wise pretraining via vector regression tops 8 medical tasks","feed_subtitle":"A displacement-field loss measures feature dispersion, preserving correlations that binary contrast destroys.","key_machinery":"The load-bearing object is the displacement vector field (DVF): a dense field whose vector at each pixel points from its position in one augmented view to the corresponding position in the other view, generated for free by a random affine spatial transformation. Inside the mapping function $V$, the vector embedding unit (VEU) computes a normalized similarity (scaled dot-product attention) between a center feature and features in an $N\\times N$ receptive field, then multiplies that distance map by a fixed vector template matrix $V^{N\\times N}$ whose entries are the coordinate offsets from the center; the weighted sum is the predicted displacement vector. Multi-vector integration (MVI) splits features into groups and averages the resulting vectors to handle ambiguous correspondences, and vector pyramid aggregation (VPA) chains these units across feature scales so the final DVF fuses global and local correspondences. The theoretical work of the key machinery is the equivalence chain showing that if the true DVF vector is representable as a weighted sum of template vectors with weights equal to normalized feature similarities, then minimizing the vector regression loss is equivalent to constraining the aggregated feature distances.","core_discovery":"The central claim is that standard binary contrastive learning, which treats every pixel pair as either positive or negative, is the wrong objective for pixel-wise pretraining because its max-min dispersion breaks the continuous feature correlations that medical images rely on. The paper defines vector contrastive learning (vector CL): a mapping function $V$ takes the dense features of two views and predicts the displacement vector field $\\psi'_{ab}$ that would align them, and the pretraining loss is the regression error $|\\psi_{ab} - \\psi'_{ab}|$ against the ground-truth DVF generated by a random affine spatial transformation. Because the DVF encodes real coordinate displacements, the mapping from feature distances to vectors is trained to be consistent with geometric correspondence; the paper shows an optimization-flow chain from distance modeling $|\\alpha_i - d'_i|$ to vector regression $|v - \\sum_i V_i d'_i|$, and argues this yields quantitatively controlled dispersion. The COVER architecture instantiates this with three components: self-vector regression (SeVR) for annotation-free ground truth, a mixture of vectors (MoV) with a vector template and multi-vector integration to handle spatial continuity and correspondence ambiguity, and vector pyramid aggregation (VPA) to capture multiscale correspondences cheaply. The paper reports that this pretraining outperforms 17 baselines on 8 tasks spanning 2D and 3D and X-ray, fundus, MR, and CT, while also improving convergence, data efficiency, and feature clustering.","pith_inferences":["Beyond the paper: if the vector-regression-to-distance equivalence is as tight as claimed, the same DVF-supervision recipe should transfer to non-medical dense tasks, for example pretraining on video or multi-view images where geometric transformations are known.","Beyond the paper: the fixed vector template could be replaced by a learnable or multi-resolution template; the paper's own ablation shows that a 7x7 template works best, which hints that the template size is a meaningful inductive bias rather than a free parameter.","Beyond the paper: a direct test of the central equivalence would be to probe the learned embedding with transformations outside the affine family, such as nonlinear deformations; if the model still quantifies distances correctly, the mechanism generalizes, and if not, the DVF supervision is learning affine-specific shortcuts.","Beyond the paper: because the method only needs a random spatial transformation, it could be pretrained jointly on heterogeneous medical datasets with shared weights, moving toward a single model that serves many organs without requiring aligned data."],"forward_implications":["Pixel-wise self-supervised pretraining can be built entirely from random geometric transformations, with no paired images or annotations, so the same recipe applies to any medical modality that can be spatially transformed.","Because dispersion is quantified rather than binary, downstream segmentation and classification tasks should inherit feature correlations that survive scenes inconsistent with the pretraining data.","The same pretrained features support smaller fine-tuning budgets: the paper reports strong downstream performance with fewer fine-tuned parameters, less fine-tuning data, and fewer pretraining iterations than the compared methods.","No separate distance-labeling stage is needed; the ground-truth distances are effectively supplied by the DVF, which sidesteps the task-dependent-distance problem that blocks direct distance modeling."],"supporting_citations":[{"why":"Shows prior vector-regression self-supervised pretraining in medical images; COVER extends it by formalizing the distance-to-vector mapping.","marker":"[26]"},{"why":"Companion vector-regression baseline that predicts dense correspondence; COVER fills its missing mapping function.","marker":"[29]"},{"why":"Provides the InfoNCE-style binary contrast objective that vector CL replaces.","marker":"[50]"},{"why":"Establishes the alignment and uniformity tension used to define the over-dispersion problem.","marker":"[74]"},{"why":"Dense binary contrastive baseline whose over-dispersion COVER targets.","marker":"[76]"},{"why":"Pixel-level consistency baseline that suffers the artificial division and over-dispersion COVER avoids.","marker":"[80]"},{"why":"Supplies the distance-metric-learning ideal that vector CL approximates without task-dependent ground-truth distances.","marker":"[82]"},{"why":"Grounds distance modeling as a principled way to preserve feature correlations, motivating the vector reformulation.","marker":"[63]"},{"why":"Chest X-ray pretraining corpus for the 2D evaluation; the largest unlabeled dataset used.","marker":"[75]"},{"why":"Brain MR pretraining corpus for the 3D evaluation; provides self-supervised data for cross-scene transfer tests.","marker":"[45]"}],"fun_headline_variants":["Vector regression beats binary contrast for pixel-wise pretraining","COVER: vector CL preserves pixel correlations, wins 8 tasks","Pixel-wise pretraining rethought: vector loss over binary","Displacement vectors fix contrastive learning's over-dispersion","Medical imaging: vector contrastive learning outperforms on 8 benchmarks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole distance-modeling argument rests on the assumption that the true displacement vector at a pixel can be expressed as a weighted sum of the fixed template vectors, with weights equal to the normalized feature similarities; if a transformation produces a displacement that the template cannot represent, the vector-regression loss can shrink without ever constraining the embedding distances to carry the intended semantics.","fun_headline_variants_meta":{"raw":{"variants":["Vector regression beats binary contrast for pixel-wise pretraining","COVER: vector CL preserves pixel correlations, wins 8 tasks","Pixel-wise pretraining rethought: vector loss over binary","Displacement vectors fix contrastive learning's over-dispersion","Medical imaging: vector contrastive learning outperforms on 8 benchmarks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000317,"raw_usage":{"total_tokens":1836,"prompt_tokens":1034,"completion_tokens":802,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":650,"completion_tokens_details":{"reasoning_tokens":717}},"tokens_in":650,"tokens_out":802,"duration_ms":7829,"temperature":1.0,"reasoning_tokens":717,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T22:40:05.853929+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Observe the correlation between learned feature distances and true spatial displacements on transformations outside the template's range, such as affine shifts larger than the template size or elastic deformations. If the Pearson correlation stays high, the distance-modeling equivalence generalizes; if it drops while the vector regression loss remains low, the model is fitting vectors without actually modeling distances, falsifying the central claim.","supporting_citations":[{"cited_title":"Geometric visual similarity learning in 3d medical image self-supervised pre- training","cited_arxiv_id":null,"evidence_quote":"Shows prior vector-regression self-supervised pretraining in medical images; COVER extends it by formalizing the distance-to-vector mapping."},{"cited_title":"Homeomorphism prior for false posi- tive and negative problem in medical image dense contrastive representation learning, 2025","cited_arxiv_id":null,"evidence_quote":"Companion vector-regression baseline that predicts dense correspondence; COVER fills its missing mapping function."},{"cited_title":"Understanding contrastive representation learning through alignment and uniformity on the hypersphere","cited_arxiv_id":null,"evidence_quote":"Establishes the alignment and uniformity tension used to define the over-dispersion problem."},{"cited_title":"Densecl: A simple framework for self-supervised dense visual pre-training.Visual Informatics, 2022","cited_arxiv_id":null,"evidence_quote":"Dense binary contrastive baseline whose over-dispersion COVER targets."},{"cited_title":"Propagate yourself: Exploring pixel-level consistency for unsupervised visual representation learning","cited_arxiv_id":null,"evidence_quote":"Pixel-level consistency baseline that suffers the artificial division and over-dispersion COVER avoids."},{"cited_title":"Distance metric learning: A com- prehensive survey.Michigan State Universiy, 2(2):4, 2006","cited_arxiv_id":null,"evidence_quote":"Supplies the distance-metric-learning ideal that vector CL approximates without task-dependent ground-truth distances."},{"cited_title":"A tutorial on distance metric learning: Mathematical founda- tions, algorithms, experimental analysis, prospects and chal- lenges.Neurocomputing, 425:300–322, 2021","cited_arxiv_id":null,"evidence_quote":"Grounds distance modeling as a principled way to preserve feature correlations, motivating the vector reformulation."},{"cited_title":"The parkin- son progression marker initiative (ppmi).Progress in neuro- biology, 95(4):629–635, 2011","cited_arxiv_id":null,"evidence_quote":"Brain MR pretraining corpus for the 3D evaluation; provides self-supervised data for cross-scene transfer tests."}],"review_version":1}