acceptodds
Under review as a conference paper at ICLR 2027

Zentropy Semantic Configuration Fields: A Cross-Modal Framework for Analyzing Semantic Organization in Learned Representations

Abstract

Representation analysis largely characterizes learned embeddings through feature similarity, linear separability, or pairwise geometry, leaving local semantic organization, i.e., how labels are composed around a point and how that composition changes with depth and neighborhood scale, without a unified measurement language. We introduce Zentropy Semantic Configuration Fields (SCF): a smoothed local class-posterior entropy, resolved as a field over network depth and neighborhood granularity , . The statistic itself is a Nadaraya-Watson label posterior followed by Shannon entropy; the contribution is the field, its descriptor calculus, and a permutation-null calibration. Controlled composition experiments show the estimator closely tracks true local semantic entropy at fixed geometry (, Pearson ), and four-quadrant experiments show it is complementary to distance rather than a reparameterization. Instantiating SCF across five modalities including vision, audio, time series, text, physiological signals reveals structured layer-by-granularity organization that often continues to evolve after linear probing has largely saturated, predominantly at fine granularities. A morphology benchmark shows that systems occupy clearly distinct field morphologies, with descriptors that are non-redundant and largely stable under seed, layer, bandwidth, and metric perturbations. A permutation-null calibration separates label-driven organization from a systematic label-free concentration drift, and reveals substantially stronger calibrated organization under label smoothing in a matched training comparison; in low-drift domains the observed organization is entirely label-driven. SCF thus provides a common analytical language for comparing how representation-learning systems organize semantic information.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
representation analysis, Zentropy, semantic configuration, multi-granularity, interpretability

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_vfNpOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes Semantic Configuration Fields (SCF), a framework for analyzing how local label composition changes across network depth and neighborhood scale. At each layer and granularity, the method estimates a distance-weighted class posterior and computes its Shannon entropy, producing a layer-by-granularity field. The paper introduces descriptors (morphology) of this field, including depth organization slope, reconfiguration depth, and organization magnitude, together with an effective-neighbor validity gate and label-permutation calibration. Controlled synthetic experiments verify sensitivity to label composition at fixed geometry (Fig. 2). Experiments across five modalities illustrate that local label organization can continue changing after linear-probe accuracy largely saturates (Fig. 4). A four-seed benchmark, primarily comparing learning objectives with a shared ResNet-18 backbone on CIFAR-100, identifies different field profiles (Table 3). The permutation control further shows that some apparent morphology differences arise from geometric concentration rather than label-driven organization.

Critical strengths

A clear measurement framework for examining local label organization across depth and scale

The paper asks a well-defined question that linear-probe accuracy and representation similarity methods does not directly answer: how are labels distributed within local neighborhoods, and how does this distribution change across layers and neighborhood sizes? The field construction makes these two axes explicit rather than selecting a single layer or granularity. The audio example in Section 6, where entropy decreases at fine granularities but increases at coarse granularities, illustrates why examining multiple scales can matter. The experiments across five modalities also provide useful evidence that the construction applies beyond one architecture or data type.

The permutation-null analysis identifies an important confound in interpreting local entropy

Section 7.4 and C.3 shows that permuting the reference labels still produces systematic depth trends: deeper layers reduce the effective number of neighbors (via clustering), allowing chance label imbalances to lower entropy even with random labels. This substantially affects the morphology descriptors: 507 of 560 null descriptor vectors fall within the real regime clusters. These results demonstrate why raw entropy trends alone are insufficient evidence of label-driven organization. Subtracting the permutation-null mean provides a useful correction, and verify that every benchmarked system exhibits organization beyond its null baseline. This analysis strengthens the measurement framework by identifying a concrete confound and showing which findings persist after correcting for it.

Critical weaknesses

The additional insight provided by SCF over established multiscale neighborhood statistics remains insufficiently demonstrated

The paper explicitly acknowledges that its estimator is a Nadaraya–Watson posterior followed by Shannon entropy, so the contribution depends primarily on the field construction, descriptors, and calibration. However, Appendix C.4 reports that established local statistics separate the coarse regimes equally well, and that a multiscale kNN-entropy baseline reproduces the depth profiles of well-organized systems with correlation 1.00.

The proposed distinctions do not yet establish a clear advantage over this closest baseline. Sensitivity to minority-class composition distinguishes entropy from maximum purity, but is also available to kNN label entropy. Similarly, a joint distance–entropy signature and permutation calibration could be applied to that baseline. The synthetic composition experiments validate the estimator’s intended behavior, but the paper does not demonstrate a substantive representation-analysis conclusion that requires SCF rather than an established local statistic evaluated across layers and scales. This limits the significance of the proposed framework beyond organizing existing measurements into a common protocol.

The evidence for distinct, reproducible semantic morphologies needs to be reconciled with calibration and the reported descriptor values

The main benchmark mentions depth reversal and cross-granularity trend. However, Section 7.4 shows that the deep reversals disappear after calibration, while the calibrated cross-granularity trend becomes identical across CIFAR-100 systems. The paper acknowledges this distinction, but much of the morphology matrix and reliability analysis concerns the raw field. Consequently, it remains unclear how much of the claimed morphological diversity persists as label-driven structure beyond differences in organization magnitude.

There are also direct inconsistencies in the reconfiguration-depth evidence. Section 8 states that CE and SupCon differ sharply in reconfiguration depth, and Appendix C.5 describes SupCon as sharing an early-reconfiguration pattern. Yet Table 5 reports (L^*=0.60) for every CE and SupCon seed, as well as every SimCLR and VICReg seed. These discrepancies affect the argument that the descriptors identify complementary, system-specific organization patterns. A complete calibrated descriptor matrix and a consistent account of reconfiguration depth are needed to assess that central claim.

Questions for authors
  1. Additional value over multiscale kNN entropy. Under the same validity and permutation-calibration protocol, which substantive findings does SCF reveal that multiscale kNN label entropy does not? Can the authors compare the calibrated descriptors and the joint distance–signature analysis directly?

  2. Morphology after calibration. Which system differences in reconfiguration depth, curvature, and cross-granularity behavior remain reproducible after null correction? A calibrated counterpart to Table 3 would help distinguish semantic morphology from the raw-field concentration effects.

  3. Reconfiguration-depth discrepancies. How should the claims about early reconfiguration in SupCon and several self-supervised systems be reconciled with Table 5, where these systems consistently have (L∗=0.60L^*=0.60)? Do the text and table refer to different fields or descriptor definitions?

1 comment

W1 / Q1. Direct comparison with multiscale kNN

Thank you for requesting a direct comparison of the calibrated descriptors and the joint distance–signature analysis against multiscale kNN.

We added this experiment using identical four-quadrant data, splits, classifiers, and paired resamples (App. E.1). The standalone SCF signature achieves 0.8840.884 accuracy, compared with 0.4850.485 for multiscale kNN and 0.8500.850 for support-matched kNN.

Importantly, SCF retains its advantage after geometric distance is included:

(d,ΓSCF)(d,\Gamma_{\mathrm{SCF}}): 0.9130.913

(d,ΓkNN)(d,\Gamma_{\mathrm{kNN}}): 0.9060.906

(d,ΓSM)(d,\Gamma_{\mathrm{SM}}): 0.9090.909

The paired gains are +0.0072+0.0072 (95% CI [0.0049,0.0093][0.0049,0.0093]) over multiscale kNN and +0.0035+0.0035 ([0.0013,0.0057][0.0013,0.0057]) over support-matched kNN.

The estimator analysis explains these differences. Support matching recovers much of the calibrated field structure (r=0.987r=0.987), while continuous weighting contributes additional composition resolution. At identical per-query support, Gaussian and Epanechnikov weighting improve posterior KL relative to uniform voting by 0.1380.138 and 0.1270.127 nats in the primary synthetic comparison. SCF also achieves lower geometric leakage and posterior error in the factorial experiment (App. E.9–E.10).

These results demonstrate additional analytical information at the signature, joint-analysis, and calibrated-field levels.

W2 / Q2. What morphology survives calibration?

Thank you for requesting a complete evaluation of morphology after removing the geometric null baseline.

We now report all eight systems' calibrated descriptors in Tables 2 and 18. The strongest differences concern final-layer organization, depth accumulation, and curvature.

Across CIFAR-100 systems, OlastO_{\mathrm{last}} ranges from 0.0910.091 for DINO-lite to 0.3550.355 for CE-LS, while ΔZcal\Delta Z_{\mathrm{cal}} ranges from 0.0230.023 to 0.2860.286. Curvature also separates systems with similar directional trends: CE has 0.698±0.0050.698\pm0.005, SupCon 0.492±0.0410.492\pm0.041, and VICReg 0.294±0.0390.294\pm0.039.

Calibration additionally changes system comparisons: the raw depth change is larger for VICReg than SimCLR, while their calibrated ordering is reversed (Fig. 2). The calibrated CIFAR fields exhibit predominantly shared directional trends but distinct organization magnitudes and trajectories. ESC-50 provides a complementary cross-modal result, with a significant opposing fine/coarse interaction after calibration (T=0.591T=0.591 nats, family-wise p=0.0025p=0.0025).

Together, these findings establish reproducible differences in the strength, accumulation, and shape of target-associated organization.

W2 / Q3. Reconfiguration-depth discrepancy

Thank you for pointing out the discrepancy between the qualitative reconfiguration description and the reported L∗L^* values.

We have corrected the interpretation. The hard reconfiguration locator gives L∗=0.60L^*=0.60 for CE and SupCon in every seed (Table 6), indicating a shared strongest transition location. Their calibrated fields nevertheless differ in final-layer organization (0.3300.330 vs 0.2850.285), depth change (0.2610.261 vs 0.2140.214), and curvature (0.6980.698 vs 0.4920.492). These measurements describe different aspects of the same depth trajectory.

We also introduced a continuous localization readout (App. C.4). On synthetic fields with known transition locations, its localization MAE decreases from 0.1030.103 for hard argmax to 0.0720.072–0.0740.074 for the soft alternatives. This provides finer localization while retaining the calibrated magnitude and curvature descriptors as the principal comparisons.

Reviewer_tHEfOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper computes the nadaraya kernel label posterior around every test point keeping the train as a reference set. It uses entropy and sweeps this over layers and badwidth tau to get a 2d field. The authors define descriptors of the field like depth slope, reconfig depth, cross granularity trend etc. They add a label permutation null to try and remove label free drift. The experiments include cover synthetic checks and 5 modalities also a cifar100 benchmarks. Around 8 setups and 4 seeds are used for training.

Critical strengths

The synthetic tests are clearly done and discussed

There is a variety in modalities and experiments - audio, time series, text, and sleep EEG are done under the protocol.

The seed table in the appendix is big and dense and a reasonable number of experiments have been reported.

Critical weaknesses

The main weakness I found was in the novelty and usability of the paper; this paper feels like a combination of many existing proven papers and methods.

Some proofs felt forced and complex; it was either a presentation issue or they were unnecessarily complicated.

There is no null calibration in the cross modal claim.

There were a lot of 1.00 values in the tables, such as in C4, and thus it remains unclear whether the field that is the paper's main contribution actually helps beyond the layerwise entropy analysis using KNN.

M is close to -1 in almost every comparison shown; this raises concerns about morphological language.

I am having doubts in trainingof the SSL models, are they fully trained? Why does label smoothing double the calibrated organization?

Questions for authors

The claim that descriptors are decoupled from performance, I feel, is not proven properly, and you need to prove that qualitatively or quantitatively. Could I please see a result around that?

Could a null calbration on the cross model claim be added?

I am having doubts in trainingof the SSL models, are they fully trained? Why does label smoothing double the calibrated organization? Could you please confirm that the models were trained properly and the loss curves are following the trend?

There is some discrepancy around the numbers, are there 28 system seed units and 560 nulls like the text says or what the table 3 shows 8 systems/32 units, could you please clarify this?

5 comments

W1–W2. Novelty, usability, and theory

Thank you for encouraging us to clarify SCF's analytical contribution and theoretical motivation.

SCF measures target organization jointly across depth and scale, calibrated against the geometry-conditioned permutation null. This changes conclusions: VICReg has a larger raw depth change than SimCLR (0.0880.088 vs 0.0740.074), but SimCLR has a larger calibrated change (0.0790.079 vs 0.0530.053; Fig. 2). In ESC-50, SCF also reveals opposing fine/coarse trajectories absent from the tested fixed-kk summaries (Fig. 3). The theory supports these findings: the Dirichlet floor defines normalization, effective support controls estimation fluctuations, and the permutation-bias analysis explains null drift. We have streamlined the exposition and retained proofs in App. A.

W3 / Q2. Cross-modal calibration

Thank you for requesting direct null validation across modalities.

We added 80 label permutations to each of six audio, time-series, and text systems. All six headline effects exceed their null references (empirical p=1/81p=1/81, BH-adjusted q=0.0123q=0.0123). For ESC-50, a 400-permutation max-statistic test across nine bandwidth pairs confirms the fine/coarse interaction (T=0.591T=0.591 nats, family-wise p=0.0025p=0.0025 in both seeds; App. E.7). Fine-scale organization increases and coarse-scale organization decreases after calibration. Support-matched kNN recovers the interaction; the tested fixed-kk variants do not.

W4–W5. Regime scores and morphology

Thank you for asking which descriptors remain informative after calibration.

Tables 2 and 18 now report the full calibrated matrix. Organization magnitude, depth change, and curvature retain strong discrimination: the CIFAR-100 between-system variance ratios for the first two are 420420 and 391391; calibrated curvature differs between CE (0.698±0.0050.698\pm0.005) and SupCon (0.492±0.0410.492\pm0.041). Ordinal markers serve as coarser descriptors.

To establish analytical value beyond the descriptive 1.001.00 regime scores, we added controlled tests against analytic ground truth and matched estimators. In the four-quadrant task, SCF's signature achieves 0.8840.884 accuracy versus 0.4850.485 for multiscale kNN. With distance included, the paired gain is +0.0072+0.0072 (95% CI [0.0049,0.0093][0.0049,0.0093]; App. E.1). These results demonstrate the field's additional analytical resolution.

W6 / Q3. SSL training and label smoothing

Thank you for suggesting training-duration and accuracy controls.

We audited SSL learning curves, trained the separate utility bank for 100 epochs, and extended VICReg to 200 epochs. Its probe accuracy rises from ∼0.442\sim0.442 to 0.481/0.4690.481/0.469, while calibrated organization changes only from 0.240/0.2540.240/0.254 to 0.258/0.2710.258/0.271; the morphology ordering persists (App. C.1).

Under the matched 35-epoch protocol, CE-LS increases OlastO_{\mathrm{last}} from 0.1970.197 (CE*) to 0.3550.355. We then compared models at nearly equal probe accuracy: two pairs within 1.5 percentage points retain organization gains of +0.108+0.108 and +0.115+0.115 (Table 11). The label-smoothing contrast therefore remains visible at comparable predictive quality.

Q1. Morphology versus predictive performance

Thank you for asking what SCF adds beyond accuracy.

Accuracy evaluates readout quality; SCF characterizes local target organization across depth and scale. The matched-accuracy pairs above have distinct calibrated fields, and organization continues to evolve after probe accuracy plateaus. A configuration-adjusted analysis finds a partial R2=0.35R^2=0.35 for the association between accuracy and OlastO_{\mathrm{last}}, quantifying their relationship while the field reveals additional depth- and scale-resolved structure (Secs. 6, 9).

Q4. Replication counts

Thank you for flagging the cohort-count ambiguity.

The complete morphology benchmark has 8×4=328\times4=32 system–seed units; its CIFAR-only subset has 7×4=287\times4=28 (including CE*). A separate earlier null-overlay cohort has 7×4=287\times4=28 units (excluding CE*) and 7×80=5607\times80=560 stored null descriptors. The utility bank contains 28 independently trained models. Appendix B.3 now distinguishes these cohorts explicitly.

Thank you again for your detailed review. We would like to briefly highlight the main additions addressing your concerns about SCF's novelty and analytical value.

The revision now includes the complete calibrated morphology matrix, permutation-calibrated cross-modal results, extended SSL training controls, and direct matched comparisons against multiscale kNN. In particular, calibration changes the VICReg/SimCLR depth-organization ordering, and the audio fine/coarse interaction remains significant after family-wise correction (p=0.0025p=0.0025).

We hope these results clarify the contribution of SCF. We would greatly appreciate your assessment of whether the revised evidence addresses your principal concerns, and we would be happy to clarify any remaining questions.

Reviewer_tHEfOpenReview ↗

Response to the rebuttal.

Thank you for addressing my concerns. The revision is clearer and more readable, and the cross-modal calibration, the cohort clarification, and the softening of claims address several concerns I initially had. I would like to retain my original score; I feel there could be more evaluations and additional experiments done to prove the method's advantage over kNN statistics. Currently, it is shown to perform well on toy datasets with limited predictive gains on real data, particularly over the probe accuracy alone. Also, the performance of support matched kNN is almost equal to SCF. To clearly judge the significance of the method, its advantage over these baselines should be stronger and shown on more real data.

Reviewer_WPaoOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This work defines a metric which is the entropy of a gaussian kerneled soft-kNN classifier with laplace/Dirichlet smoothing. They sweep the kernel bandwidth across network layers to characterize how class representations within a discrete class is organized across spatial scales and depth. While the terminology is overly grandiose and simple kk-NN baselines recover the same broad regimes, the authors' use of effective-sample size gating and label-permutation nulls is a solid contribution that separates true semantic organization from geometric clustering artifacts.

Critical strengths
  1. The permutation-null calibration (Sec. 7.4) is a reasonable way to separate true label organization from geometric clustering artifacts. Shuffling the labels shows that tighter clustering in deeper layers artificially lowers entropy, which explains away the apparent deep-layer reversals in SimCLR and VICReg (line 444). Good statistical hygiene, even if just standard permutation test.

  2. The ablations and robustness checks are very thorough and honest. The authors openly show that simple k-NN entropy and gini impurity separate the broad regimes just as well line 1002, while verifying that their descriptors hold up under whitening, cosine distance, and layer perturbations in line 461.

  3. They find that neighborhoods keep purifying long after probe accuracy saturates (fig 4), and in audio, fine and coarse scales actually move in opposite directions with depth. The probe accuracy saturation problem is something we observe in our own research as well.

Critical weaknesses
  1. The method is a bit "over-packaged" when simple baselines do almost the same job. Aside from terms like "Zentropy" and "morphology calculus," this is just kernel label entropy swept over a bandwidth grid, and Appendix C.4 shows that basic kk-NN entropy (k=10k=10) separates the training regimes just as well without needing bandwidth sweeps or 80 permutation runs to fix variable neighborhood sizes. (Although I appreciate the honesty!)

  2. It omits comparisons to intrinsic dimensionality estimators like TwoNN (Elena Facco). The authors argue in Section 7.4 that deep-layer geometric tightening shrinks effective neighborhood sizes and drives their null drift, so comparing against a hyperparameter-free geometric probe like TwoNN would have directly tested whether intrinsic dimension contraction explains this drift.

  3. The framework only works for discrete classification with moderate class counts. Because it relies on a categorical posterior over KK classes, it cannot handle continuous regression, multi-label tasks, or unannotated pretraining data, and its finite-sample bias worsens as KK grows large relative to the neighborhood size.

Questions for authors
  1. Why use a fixed-bandwidth kernel that requires an 80-run permutation null to fix variable neighborhood sizes, rather than a multiscale kk-NN entropy estimator that fixes the neighbor count kk directly?

  2. How does the permutation-null drift across layers correlate with established intrinsic dimensionality estimators like TwoNN (Facco et al., 2017; Ansuini et al., 2019)?

  3. Why does the Dirichlet smoothing parameter α\alpha switch from a fixed constant (0.050.05) to a KK-scaled value (0.25/K0.25/K) in the vision benchmark?

  4. Can the authors clarify the discrepancy in the descriptor PCA explained variances reported between Appendix C.3 (55.2% and 25.4%) and Appendix C.7 (62.8% and 22.9%)?

  5. What about tasks that aren't classification? Any thoughts?

5 comments

W1 / Q1. Fixed bandwidth and comparison with kNN

Thank you for highlighting the need to distinguish SCF from conventional multiscale neighborhood statistics.

The distinction is what the neighborhood scale represents. Fixed-kk voting changes its radius across queries, whereas SCF holds a normalized metric bandwidth fixed and measures both label composition and the effective support available at that scale. This makes geometry-dependent changes in the local entropy baseline measurable. In our 28-run comparison, the permutation-null depth slopes are +0.006+0.006 for fixed-kk and −0.70-0.70 for SCF (Sec. 6.4; App. E.9).

We further isolated the contribution of distance weighting. At identical per-query effective support, Gaussian and Epanechnikov weighting improve posterior KL over uniform voting by 0.1380.138 and 0.1270.127 nats, respectively, in the primary synthetic comparison.

The four-quadrant experiment provides a direct analytical comparison. The SCF signature achieves 0.8840.884 accuracy versus 0.4850.485 for multiscale kNN; after adding distance, (d,ΓSCF)(d,\Gamma_{\mathrm{SCF}}) achieves 0.9130.913 versus 0.9060.906 for (d,ΓkNN)(d,\Gamma_{\mathrm{kNN}}), with a paired improvement of +0.0072+0.0072 (95% CI [0.0049,0.0093][0.0049,0.0093]; App. E.1).

These results establish SCF's additional value in local-composition estimation and representation analysis. Calibration reuses cached representations and requires no additional model training.

W2 / Q2. Intrinsic dimensionality and null drift

Thank you for suggesting intrinsic dimensionality as an alternative explanation for the observed null drift.

We evaluated TwoNN intrinsic dimension on the 28-run model bank. Its association with null drift is weak (Spearman ρ=0.08\rho=0.08, cluster-bootstrap CI [−0.11,0.22][-0.11,0.22]), including after controlling for effective support (partial ρ=0.07\rho=0.07). In contrast, within-run null entropy tracks effective support more closely (mean ρ=0.64\rho=0.64).

Our finite-population analysis identifies the leading unsmoothed entropy bias:

\approx -\frac{K-1}{2N_{\mathrm{eff}}} \frac{N-N_{\mathrm{eff}}}{N-1}.$$ Query-level experiments establish the quantitative operating regime: after smoothing correction, relative bias error is $4.6%$ for $N_{\mathrm{eff}}/K\in[5,20)$. The combined analytical and experimental results identify finite effective support as the direct statistical mechanism behind the null drift (App. E.6, E.8). ### W3 / Q5. Extension beyond classification Thank you for encouraging us to examine the broader applicability of the measurement framework. We implemented C-SCF, replacing categorical posteriors with local conditional-density estimates and calibrating their divergence from the target marginal. The extension is evaluated on analytic synthetic regression, California Housing with five training seeds and spatial block splitting, and subject-held-out PAMAP2 heart-rate regression. All five housing seeds exhibit increasing organization across depth. PAMAP2 further demonstrates that preserving subject grouping and temporal dependence materially changes the permutation reference (App. E.12). SCF also applies to encoders pretrained without annotations: our frozen CLIP, DINOv2, and MAE experiments measure organization using labeled evaluation references. For multilabel targets, separate calibrated binary fields provide label-wise measurements, while joint target states provide a natural multivariate formulation. The large-class-count setting is examined through the explicit $N_{\mathrm{eff}}/K$ analysis and experiments with $K\in{5,20,50,100}$. ### Q3. Dirichlet smoothing Thank you for pointing out the need to clarify the smoothing parameterization. The relevant quantity is the total Dirichlet pseudomass, $\gamma=K\alpha$. At $K=100$, setting $\alpha=0.05$ produces $\gamma=5$, which substantially flattens the estimated posterior. We use $\gamma=0.25$ to maintain comparable regularization across class counts. A sensitivity sweep over $\gamma\in{0.05,0.25,1,5}$ preserves the CE–SimCLR ordering while changing absolute entropy amplitudes as expected (App. E.17). ### Q4. PCA explained variance Thank you for catching the inconsistent PCA figures. We recomputed the descriptor PCA using a consistent cohort and normalization. The verified values are 55.2% for PC1 and 25.4% for PC2, totaling 80.6%. These values are now consistent across the manuscript, Figure 16, and Appendix E.18.

Thanks! Could you clarify your "C-SCF" experiment? Is this for a univariate target regression task?

I continue to recommend for acceptance.

Thank you for your follow-up question and continued support.

Yes, both real-world C-SCF experiments use univariate continuous regression targets: median house value in California Housing and heart rate in PAMAP2.

We also evaluate multivariate continuous targets in controlled synthetic experiments with target dimensions dy∈{1,2,5}d_y\in\{1,2,5\}. These experiments examine how target dimensionality affects effective support and conditional-density estimation.

Thus, our real-world experiments validate C-SCF on univariate regression tasks, while the synthetic experiments additionally cover multivariate targets (Appendix E.12).

Comments on the paper

1 comment

We thank the reviewers for their detailed assessments and constructive questions. SCF measures how local target composition develops jointly across representation depth and neighborhood scale. Its central contribution is a geometry-calibrated measurement framework that combines continuous local composition estimates, effective-support accounting, and a shared label-permutation reference.

The expanded evaluation provides concrete evidence of the analytical value of this construction. Calibration changes a comparison between learning systems: VICReg has the larger raw depth change, whereas SimCLR has the larger calibrated depth change. Resolving both depth and scale exposes complementary organization: the ESC-50 example shows increasing organization at fine scales and decreasing organization at coarse scales after calibration. Direct comparisons establish an estimator-level contribution: continuous weighting improves posterior estimation over uniform support-matched voting in the tested moderate-to-large-class settings, and the four-quadrant experiment now compares SCF and kNN signatures directly, both alone and jointly with distance.

The controlled geometry–semantics experiment reports a 13.8-fold reduction in its geometry-induced variation measure while retaining monotonic sensitivity to semantic strength. The continuous-target extension further applies the same measurement principle to synthetic regression, California Housing, and subject-held-out physiological regression. Together, these results connect the statistical construction to specific questions about learned representations: how much target-associated organization remains beyond a geometric reference, how that organization changes across depth, and which neighborhood scales carry the change.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.