Low-label transfer is often decided by a practical question: after a representation looks useful, how many labels are still needed to learn the downstream boundary? Transferability and probe-quality scores characterize global alignment, likelihood, margins, and calibration. Boundary-near target mass complements these diagnostics by measuring the ambiguous regions where scarce labels are most costly. We propose effective boundary complexity, a label-budget diagnostic for frozen representations and linear probes. It measures mass in a thin strip around the learned boundary and normalizes it by class balance, so dense ambiguous regions and predicted-class imbalance are both counted. We prove that, under a smooth representation map, density change and tangential surface change cancel at first order, leaving boundary-normal stretch or compression as the governing factor. This yields a cached-feature estimator that fits a probe on one split and counts independent held-out features near its boundary. Experiments support the diagnostic at three levels. Controlled transformations match the predicted law across 74 maps with 0.38% median relative error. On real frozen features, normal-direction changes move both measured boundary mass and label demand, while tangent controls preserve the first-order coefficient. A differentiable surrogate reduces measured complexity by 16-36% at matched top-1 utility. Finally, CIFAR-100 and Tiny-ImageNet experiments support boundary mass as a second-stage diagnostic for near-tied representations. Five-way experiments show strong OvR-interface backbone-ranking agreement (Spearman 0.872), while acquisition experiments show high correlations between initial boundary mass and realized label budgets across three sampling policies (0.779-0.830). Code is available at https://github.com/qqaqq-code/boundary_mass.
Figures & tables
Figure 1: Operational label-budget evaluation for frozen representations. For a fixed downstream task, different feature maps place different probability mass inside the learned boundary tube. Lower effective boundary complexity means fewer boundary-near examples for the low-label learner to resolve.
Figure 2: Boundary mass and the pushforward mechanism. A boundary tube measures probability mass near the learned decision boundary. Under a smooth feature map T , the pushed density and tangential surface Jacobian cancel, leaving the inverse-transpose normal factor ∥DT(x)−⊤n(x)∥ in the transformed boundary mass. Boundary-normal changes affect the first-order tube coefficient; tangent-only changes preserve it.
Figure 3: Finite-sample validation of the pushforward law. Predicted and measured boundary mass agree across controlled transformations; the orientation sweep shows the boundary-normal factor ∥DT−⊤n∥ directly. Across 74 transformations, median relative error is 0.38% and mean relative error is 0.50% .
Figure 4: Normal and tangent scaling of frozen features. Full-data heads provide the fixed-bandwidth kernel score Dker ; low-label heads provide learning curves and budgets. From s=1 to s=2 , normal expansion lowers mean budget by 1.54 spc, while the tangent control changes it by +0.02 spc.
s
log(1+Dker)↓
5-spc acc. ↑
AULC ↑
budget ↓
0.5
0.129
88.34
92.072
8.47
1.0
0.072
94.50
96.066
3.62
2.0
0.020
98.06
98.323
2.08
4.0
0.006
98.92
99.023
2.00
Table 1: Normal and tangent scaling on 8 tasks and two backbones, with three label-draw seeds per budget. Left: mean normal-scaling score, accuracy at 5 spc, AULC on a 0–100 scale, and budget in spc. Right: paired changes from s=1 to s=2 . Kernel scores use full-data heads and the s=1 bandwidth.
Setting
Method
Top-1 ↑
ψ^↓
Low-label probe ↑
CIFAR-100 10 spc
standard / ψ^ -int.
53.97/53.84
201.68/169.22
–
CIFAR-100 50 spc
standard / ψ^ -int.
65.82/65.74
246.23/185.57
–
Tiny-ImageNet 10 spc
standard / ψ^ -int.
62.85/62.73
507.04/387.09
–
Tiny-ImageNet 50 spc
standard / ψ^ -int.
72.15/71.98
659.58/420.08
–
CIFAR-100 R18
standard
53.68±0.25
23.57±0.33
44.46/47.47
CIFAR-100 R18
weight decay
54.55±0.20
23.61±0.23
45.53/48.43
Table 2: Smooth surrogate intervention. Frozen-feature rows compare standard linear-head training with a ψ^ intervention at the same samples per class; their low-label utility is the Top-1 column, so the probe column is reserved for the end-to-end rows. End-to-end rows use the top-2 boundary-pressure proxy and report mean ± standard deviation over three CIFAR-100 ResNet-18 training seeds, each evaluated with three downstream probe seeds. Header arrows mark preferred directions.
Figure 5: Backbone-selection evaluation. Left: held-out boundary mass and realized label budget. Right: incremental ΔR2 from boundary mass after transferability controls. The full diagnostic set improves from R2=0.734 to 0.749 , supporting complementary second-stage selection among near-tied candidates.
Comparison
Spearman ↑
Concordance ↑
Backbone agreement ↑
Hard OvR vs. Hard Interface
0.872
90.0%
89.5%
Smooth OvR vs. Smooth Interface
0.868
89.7%
89.5%
Table 3: Multiclass estimator agreement on 40 five-way Tiny-ImageNet tasks, four frozen backbones, and five random splits with shared semantic class identities across backbones. The metrics quantify agreement in backbone rankings.
Policy
Realized budget ↓
Spearman ↑
Selection regret ↓
Random
7.05
0.779
1.08
Uncertainty
5.57
0.830
0.58
Hybrid
5.47
0.824
0.42
Table 4: Acquisition-policy sensitivity on 40 CIFAR-100 binary tasks and four frozen backbones. Initial Gaussian-kernel Dker uses a probe fit on the shared class-balanced 2-spc labeled set and diagnostic-heldout features. Budgets and selection regrets are in spc.
Appendix figures & tables39 assets
Supplementary material from the paper’s appendix.
Appendix
Check
heads
Pearson ↑
Spearman ↑
error / diff ↓
split-half log(1+Dker)
160
0.960
0.970
3.2%
hard-tube vs. kernel
160
0.981
0.982
0.026
Appendix
Table 5: Estimator reliability on the strict diagnostic-heldout split. Split-half entries summarize 100 random stratified halves. Hard-tube agreement compares log(1+Dhard) and log(1+Dker) with the same quantile width. Correlations are computed across 160 heads.
Setting
n
rank stability ↑
median shift ↓
eps0.05/tau0.025/rho1e-06
54
0.993
1.0%
eps0.05/tau0.05/rho1e-06
54
0.999
23.6%
eps0.1/tau0.025/rho1e-06
54
0.998
5.2%
eps0.1/tau0.05/rho1e-06
54
1.000
0.0%
eps0.1/tau0.1/rho1e-06
54
0.993
21.9%
eps0.2/tau0.05/rho1e-06
54
0.991
5.9%
Appendix
Table 6: Estimator sensitivity and cost. Rank stability is Spearman correlation of ψ^eff under a hyperparameter setting against the default (ε0,τ,ρ)=(0.1,0.05,10−6) . Cost is mean seconds per epoch from saved training logs.
Setting
heads
rank stability ↑
mean regret ↓
rel. regret ↓
AULC ↑
fixed ε0=0.05
160
0.671
0.21
10.7
0.975
q0.1 tube
160
0.939
0.09
4.4
0.975
q0.2 tube
160
1.000
0.10
5.0
0.977
q0.3 tube
160
0.998
0.08
4.0
0.978
margin-scaled tube
160
0.942
0.02
0.9
0.980
hard q0.2 tube
160
0.982
0.11
5.6
0.978
Appendix
Table 7: Tube-width robustness on the strict CIFAR-100 evaluation. Rank stability is Spearman correlation against the default standardized q0.2 kernel estimate. Nearby quantile widths and the hard-tube estimator preserve rankings; a globally fixed width is less stable because signed-distance scales differ across heads.
Protocol
heads
rank stability ↑
mean regret ↓
rel. regret ↓
AULC ↑
raw features
160
0.369
0.33
16.3
0.973
standardized
160
1.000
0.10
5.0
0.977
L2-normalized
160
0.683
0.05
2.4
0.979
PCA-whitened
160
0.232
0.33
16.6
0.972
Appendix
Table 8: Feature-normalization evaluation on the strict CIFAR-100 backbone selection task. The diagnostic is stable under the paper’s standardized protocol and nearby normalized variants, while raw and principal component analysis (PCA)-whitened protocols change the ranking substantially. This supports comparing absolute Dker values under a fixed protocol.
Benchmark
runs
classes
feature dim.
CIFAR-100/R50
106
100
2048
CIFAR-100/ViT-B
27
100
768
Tiny-IN/R50
45
200
2048
Appendix
Table 9: Benchmarks included in the completed proxy evaluation. The evaluation covers both CIFAR-100 and Tiny-ImageNet cached-feature settings.
Asset
Used for
Citation
CIFAR-100
binary class-pair experiments, interventions
[ Krizhevsky, 2009 ]
Tiny-ImageNet
external replication
[ Le and Yang, 2015 ]
ImageNet / ILSVRC
pretraining source for frozen backbones
[ Deng et al., 2009 ; Russakovsky et al., 2015 ]
ResNet-50 / ResNet-18
backbone / end-to-end evaluation
[ He et al., 2016 ]
ViT-B/16
frozen backbone
[ Dosovitskiy et al., 2021 ]
DINO-ViT-B/16
frozen backbone
[ Caron et al., 2021 ]
Appendix
Table 10: Datasets and pretrained backbones used in the main experiments, with their roles and original references.
Mode
runs
median dist. ↓
mean dist. ↓
expansion+compression ↑
psieff
118
0.417
0.516
0.0%
standard
54
0.951
0.875
0.0%
weight decay
78
0.929
0.855
0.0%
Appendix
Table 11: Boundary–Utility final distance to the empirical envelope. The intervention changes boundary geometry at comparable final utility, and this table visualizes the resulting diagnostic state. The expansion+compression column is the fraction of runs whose final point improves both boundary compression and utility expansion relative to the empirical envelope; all modes are zero under this strict joint criterion.
Full-probe evaluation
Acquisition correlation
Score
ΔR2
Top-2 regret
Random
Uncertainty
Hybrid
Dker
0.015
0.030
0.779
0.830
0.824
ψeffker
0.012
0.030
0.809
0.855
0.849
Dhard
0.023
0.081
0.814
0.860
0.855
ψeff,ε0hard
0.022
0.034
0.829
0.872
0.867
Appendix
Table 12: Kernel and tube estimates for label-budget prediction. Full-probe regressions use 160 task–backbone pairs and a conventional baseline with R2=0.734 . Top-2 regret is in spc. Acquisition columns report Spearman correlations between seed-averaged log-score and realized label budget.
Figure 6: Strict backbone-selection label-budget evaluation across 40 CIFAR-100 class pairs and four cached backbones, including DINO and SigLIP. Diagnostics are computed on held-out training features; test labels are reserved for the target learning curves, preserving the prospective selection protocol.
Baseline set
base R2↑
+ DkerR2↑
ΔR2↑
p↓
NLL + ECE
0.343
0.462
0.120
0.007
margin + effective rank
0.171
0.434
0.263
<0.001
NLL + ECE + margin + rank
0.399
0.530
0.130
<0.001
all non- Dker proxies
0.513
0.660
0.147
<0.001
Appendix
Table 13: Main CIFAR-100 four-backbone regression evaluation. R2 predicts pair-demeaned label budget from the listed non- Dker covariates, then adds log(1+Dker) . p is a paired sign-flip test for the incremental ΔR2 .
Baseline set
base R2↑
+ DkerR2↑
ΔR2↑
p↓
partial corr.
LogME + H-score
0.509
0.545
0.036
0.027
0.270/0.180
LogME + H-score + TransRate
0.538
0.589
0.050
0.004
0.330/0.258
feature transferability set
0.612
0.646
0.034
0.009
0.297/0.260
all non- Dker + transferability
0.734
0.749
0.015
0.039
0.239/0.233
Appendix
Table 14: CIFAR-100 four-backbone regression evaluation with transferability controls. Feature-transferability metrics are computed only on diagnostic-heldout training features.
Metric
label-budget corr.
selection acc. ↑
mean regret ↓
rel. regret ↓
log(1+Dker)
0.493/0.266
25.0
0.10
5.0
Dker
0.499/0.266
25.0
0.10
5.0
Gaussian-area proxy
0.240/0.036
30.0
0.38
19.0
noise flip proxy
0.207/0.157
50.0
0.84
42.1
median margin
-0.293/-0.239
22.5
0.13
6.6
held-out NLL
0.525/0.404
12.5
0.18
8.9
Appendix
Table 15: Strict single-metric backbone-selection results for label budget over ResNet-50, ViT-B/16, DINO, and SigLIP. Diagnostics are computed on held-out training features only.
Metric
label-budget corr.
LogME
−0.313/−0.418
H-score
−0.636/−0.470
TransRate
0.217/0.284
NCE-style cluster
−0.645/−0.575
Appendix
Table 16: Feature-transferability single-metric correlations with pair-demeaned label budget under the strict diagnostic protocol. These metrics use diagnostic-heldout training features and labels. Higher LogME/H-score/TransRate/NCE-style values are conventionally better, so negative correlations mean lower label demand.
Baseline set
base R2↑
+ DkerR2↑
ΔR2↑
p↓
partial corr.
PACTran-linear
0.211
0.281
0.071
0.110
0.299/0.071
feature transferability + PACTran-linear
0.625
0.648
0.024
0.059
0.251/0.257
all non- Dker + transferability + PACTran-linear
0.745
0.758
0.014
0.092
0.232/0.207
Appendix
Table 17: PACTran-linear baseline on the strict CIFAR-100 backbone-selection evaluation. This is a PAC-Bayes-style regularized linear target-head score on diagnostic-heldout features, implemented as a practical PACTran proxy.
Baseline set
base R2↑
+ DkerR2↑
ΔR2↑
p↓
partial corr.
LEEP + NCE
0.199
0.248
0.049
0.519
0.247/0.234
source-logit set
0.260
0.331
0.071
0.342
0.310/0.284
source + feature transferability
0.653
0.715
0.062
0.040
0.423/0.435
all non- Dker + source/feature
0.863
0.863
0.000
0.870
0.036/0.047
Appendix
Table 18: Source-logit transferability evaluation on the ResNet-50/ViT-B/16 subset. We downloaded/loaded the official torchvision ImageNet classifiers, forwarded CIFAR-100 training images, and computed LEEP, NCE, and source-logit LogME on diagnostic-heldout training examples. This source-head evaluation uses the two backbones with comparable 1000-way ImageNet classifiers.
Metric
label-budget corr.
LEEP
−0.433/−0.686
NCE
−0.400/−0.571
source-logit LogME
−0.228/−0.528
Appendix
Table 19: Single source-logit transferability correlations with pair-demeaned label budget on the ResNet-50/ViT-B/16 subset. Higher values are conventionally better, so negative correlations mean lower label demand.
Baseline set
base R2↑
+ DkerR2↑
ΔR2↑
p↓
partial corr.
NLL + ECE
0.595
0.657
0.062
0.072
0.392/0.231
margin + effective rank
0.377
0.649
0.272
<0.001
0.661/0.758
feature transferability
0.636
0.658
0.023
0.174
0.249/0.336
all non- Dker proxies
0.688
0.701
0.013
0.290
0.203/0.357
all non- Dker + transferability
0.701
0.709
0.008
0.388
0.166/0.311
Appendix
Table 20: Independent Tiny-ImageNet label-budget evaluation over 40 binary class pairs and four frozen backbones. ViT-B/16, DINO, and SigLIP Tiny features were extracted for this evaluation; ResNet-50 uses the existing Tiny-ImageNet feature cache. Diagnostics are computed on diagnostic-heldout training features, and the incremental gains indicate that the binary external evaluation is close to saturation.
Metric
selection acc. ↑
mean regret ↓
rel. regret ↓
mean AULC ↑
log(1+Dker)
77.5
0.01
0.5
0.985
LogME
87.5
0.00
0.0
0.985
H-score
87.5
0.61
30.5
0.982
NLL
37.5
0.08
3.8
0.982
ECE
42.5
0.33
16.4
0.980
margin
77.5
0.01
0.5
0.985
Appendix
Table 21: Tiny-ImageNet single-metric backbone-selection summary. This independent evaluation provides replication evidence: Dker has low regret, while transferability metrics such as LogME are also strong.
Setting
tasks
backbones
base regret ↓
ψ regret ↓
reduction ↑
p↓
AULC ↑
Tiny-ImageNet 5-way OvR
16
4
0.051
0.100
−0.049
0.709
0.961
Tiny-ImageNet 5-way interface
16
4
0.051
0.450
−0.398
0.999
0.947
Appendix
Table 22: Tiny-ImageNet 5-way aggregation stress test. The evaluation compares the operational OvR score with an interface-counted argmax-geometry variant under the same cached-feature protocol. The 16-task evaluation reports backbone-selection regret on a label grid where most tasks reach the target at the smallest budget.
Policy
regret ↓
reduction ↑
Δ AULC ↑
win rate ↑
Non- ψ base
0.051
0.000
0.000
0.00
OvR top-2
0.100
−0.049
0.001
0.12
OvR abstain
0.047
0.004
0.001
0.12
Interface top-2
0.450
−0.398
−0.013
0.00
Interface abstain
0.145
−0.093
−0.004
0.00
Non- ψ + OvR reg.
0.051
0.000
0.000
0.00
Appendix
Table 23: Tiny-ImageNet 5-way policy variants under saturation. Abstention and regression variants test whether ψ should act only when non-boundary diagnostics are near-tied. The 16-task evaluation reports policy regret, budget reduction, AULC change, and win rate; the OvR abstention policy yields a positive budget reduction.
Label policy
Realized budget ↓
Spearman (kernel Dker , budget) ↑
Selection regret ↓
Random
7.05
0.779
1.08
Uncertainty
5.57
0.830
0.58
Hybrid
5.47
0.824
0.42
Appendix
Table 24: Label-acquisition policy sensitivity on 40 CIFAR-100 binary tasks and four frozen backbones. A shared class-balanced 2-spc labeled set initializes the probe for Smooth boundary-mass estimation, followed by Random, Uncertainty, or Hybrid acquisition. Budgets and selection regrets are in spc.
pilot spc
policy
selection acc. ↑
mean regret ↓
rel. regret ↓
mean AULC ↑
2
choose by Dker
17.5
0.05
2.4
0.979
2
choose by margin
15.0
0.09
4.7
0.978
2
NLL/ECE
22.5
0.06
3.2
0.978
2
LogME/H-score
52.5
0.13
6.3
0.971
2
transferability set
27.5
0.05
2.6
0.977
2
combined non- Dker
30.0
0.03
1.3
0.975
Appendix
Table 25: Prospective CIFAR-100 pilot-label policy simulation. Heads use 2, 5, or 10 labels per class; diagnostics are computed on held-out training features and policies are leave-one-class-pair-out budget regressions.
pilot spc
tasks with lower budget ↑
mean budget reduction ↑
mean relative reduction ↑
2
2.5
−0.01
−0.5
5
0.0
0.00
0.0
10
0.0
−0.02
−0.9
Appendix
Table 26: Paired effect of adding Dker to the combined non- Dker pilot policy. Positive reduction means the selected backbone has lower label budget.
Pilot spc
Refit SD ↓
Backbone-order agreement ↑
Lower- Dker exact ↑
Combined+ Dker exact ↑
Budget agreement ↑
2
0.018
77.2
55.8
57.2
92.0
5
0.008
83.0
66.2
52.2
91.5
10
0.005
83.5
64.5
54.3
91.5
Appendix
Table 27: Pilot-refitting stability on 40 CIFAR-100 binary tasks and four backbones, with five pilot draws per configuration and a fixed diagnostic pool. Refit SD summarizes the across-draw standard deviation of log(1+Dker) . Exact agreement refers to the selected backbone under the lower- Dker and combined+ Dker policies; budget agreement refers to the combined+ Dker policy. Agreement columns are percentages.
Policy
selection acc. ↑
mean regret ↓
rel. regret ↓
mean AULC ↑
log(1+Dker)
25.0
0.10
5.0
0.977
LogME
22.5
0.02
0.9
0.980
H-score
25.0
0.03
1.4
0.980
TransRate
42.5
1.59
79.5
0.959
transferability set
27.5
0.05
2.6
0.977
transferability set + Dker
27.5
0.05
2.6
0.977
Appendix
Table 28: Prospective backbone-selection policy simulation. Policies are leave-one-class-pair-out regressions from the listed diagnostics to label budget; selection uses only diagnostic-heldout training features.
Figure 7: Near-tie selection evaluation. Lower Dker is most informative when non-boundary metrics produce a near tie; large conventional-score gaps define the abstention regime.
Full-shot acc. gap
comparisons
lower- Dker win/tie score ↑
mean relative saving ↑
≤1%
176
54.3
1.2
≤2%
227
54.6
2.4
Appendix
Table 29: Matched-utility pairwise evaluation under the strict diagnostic protocol. For the lower- Dker backbone, win/tie scores assign 1 , 0.5 , and 0 to lower, equal, and higher label budgets, respectively. Comparisons group backbone pairs by full-shot accuracy gap.
Metric
acc.-drop corr.
robust-selection acc. ↑
log(1+Dker)
0.222/0.036
50.0%
Gaussian-area proxy
−0.202/0.054
40.0%
noise flip proxy
0.588/−0.029
45.0%
median margin
−0.199/−0.026
52.5%
NLL
0.065/0.175
72.5%
ECE
0.157/0.165
75.0%
Appendix
Table 30: Feature-noise control evaluation. Clean accuracy, NLL, and ECE dominate robust-accuracy selection under isotropic feature noise, while Dker measures the complementary boundary-mass axis.
Backbone
Evaluation status
CLIP ViT-B/16
Pending evaluation
DINO ViT-B/16
Completed
DINOv2 Base
Pending evaluation
MAE ViT-Base
Pending evaluation
SigLIP-B/16
Completed
Appendix
Table 31: Foundation-backbone evaluation status. Completed experiments include DINO and SigLIP alongside ResNet-50 and ViT-B/16.
Comparison
Within-task Spearman ↑
Pairwise concordance ↑
Selected-backbone agreement ↑
Mean rank displacement ↓
Hard OvR vs. Hard Interface
0.872
90.0%
89.5%
0.2925
Smooth OvR vs. Hard OvR
0.988
99.0%
98.5%
0.0300
Smooth Interface vs. Hard Interface
0.992
99.3%
99.5%
0.0200
Smooth OvR vs. Smooth Interface
0.868
89.7%
89.5%
0.3025
Appendix
Table 32: Agreement among multiclass boundary-mass estimators on 40 five-way Tiny-ImageNet tasks, four frozen backbones, and five random splits with shared semantic class identities across backbones. The comparisons isolate multiclass aggregation, smoothing, and their combined effect on backbone ranking.
Figure 8: Controlled pushforward validation. Predicted ψeff and measured tube estimates agree across scalings, shears, and boundary orientations. The orientation plot validates the boundary-normal mechanism ∥DT−⊤n∥ directly.
Figure 9: Sample-complexity collapse. Different representations produce different raw learning curves. Normalizing by n/ψeff collapses the curves and recovers the expected two-dimensional exponent.
Figure 10: Grouped Boundary–Utility Plane from the frozen-feature runs. The grouping by dataset and sample regime reveals interpretable trajectories: CE, weight decay, and ψeff intervention move through utility–boundary space with different slopes and efficiency profiles. The plane is used as a coordinate system for training dynamics, complementing a theorem-level optimality claim.
Predictor
spc-resid. Pearson
spc-resid. Spearman
mixed Pearson
Runs
ψ^eff
0.200
0.228
0.369
106
log(1+ψ^eff)
0.204
0.236
0.339
106
Weight norm
0.070
0.127
0.965
106
Normalized margin
0.576
0.567
0.996
106
Feature effective rank
0.442
0.409
0.960
106
Noise flip rate
−0.183
−0.088
−0.884
106
Appendix
Table 33: Completed proxy evaluation on 106 selected linear heads on frozen ResNet-50 features for CIFAR-100. All rows are computed from the fitted heads and their feature data. The main columns use spc-residual correlations, which compare runs at a matched label budget; the mixed Pearson column is included to show how strongly raw correlations can be driven by sample budget.
Dataset
spc
Standard Top-1 / ψ^eff
ψeff -int. Top-1 / ψ^eff
Complexity drop
CIFAR-100
10
53.97%/201.68
53.84%/169.22
−16.1%
CIFAR-100
20
60.03%/215.13
59.90%/173.80
−19.2%
CIFAR-100
50
65.82%/246.23
65.74%/185.57
−24.6%
Tiny-ImageNet
10
62.85%/507.04
62.73%/387.09
−23.7%
Tiny-ImageNet
20
68.01%/558.31
67.84%/397.38
−28.8%
Tiny-ImageNet
50
72.15%/659.58
71.98%/420.08
−36.3%
Appendix
Table 34: Smooth surrogate as geometry intervention on frozen ImageNet-pretrained ResNet-50 features. The intervention moves feature-space boundary complexity strongly while preserving nearly the same utility, illustrating controllability of the measured geometry.
Figure 11: CNN training monitoring on MNIST digits 3 and 5. The curves display the feature-space boundary readout and test accuracy over training, illustrating joint changes in representation geometry and predictive performance.
Method
Test Acc. ↑
ψ^hard↓
ψ^kern↓
ECE ↓
∥W∥F
feat. norm
margin
20/50-spc probe ↑
Standard
53.68±0.25%
23.57±0.33
11.88
0.060
10.38
31.60
0.322
44.46/47.47
Weight decay
54.55±0.20%
23.61±0.23
11.77
0.055
10.03
30.21
0.361
45.53/48.43
Margin reg.
54.55±0.18%
23.82±0.52
11.80
0.046
10.48
31.42
0.371
44.96/47.96
Entropy reg.
53.98±0.13%
24.01±0.41
12.05
0.051
10.34
31.40
0.304
44.49/47.43
ψeff -reg, λ=0.24
57.09±0.42%
22.07±0.35
10.95
0.048
9.69
32.48
0.592
48.26/51.07
Appendix
Table 35: CIFAR-100 ResNet-18 end-to-end controllability check. Entries are mean ± standard deviation over three training seeds. Low-label columns are frozen linear-probe top-1 accuracies on the epoch-40 representations, averaged over three label-subsampling seeds inside each training seed.
λ
Test Acc. ↑
ψ^hard↓
ψ^kern↓
ECE ↓
20/50-spc probe ↑
0.06
56.27±0.29%
23.56±0.19
11.77
0.069
47.78/50.55
0.12
57.05±0.27%
23.22±0.55
11.59
0.057
48.09/50.63
0.24
57.09±0.42%
22.07±0.35
10.95
0.048
48.26/51.07
Appendix
Table 36: ResNet-18 λ sweep on CIFAR-100. The sweep shows a Pareto-safe region: larger boundary pressure lowers both hard and kernel tube estimates while preserving or improving test accuracy and low-label probes.
Method
Test Acc. ↑
ψ^hard↓
ψ^kern↓
ECE ↓
20/50-spc probe ↑
Standard
84.36±0.17%
1.96±0.07
1.00
0.051
81.78/82.32
Weight decay
85.08±0.23%
2.12±0.11
1.03
0.047
82.09/83.10
ψeff -reg
86.13±0.16%
1.54±0.07
0.76
0.042
83.88/84.39
Appendix
Table 37: CIFAR-10 ResNet-18 repeat. The same boundary-pressure run improves test accuracy, hard/kernel boundary estimates, expected calibration error (ECE), and low-label probes over standard and weight-decay training.