Fixed-budget frontier

Fourteen layers had the highest observed mean.

The 1.0x column keeps the parameter budget near 5M while depth changes. D=14 has the highest mean accuracy at this budget on both CIFAR tasks, with D=20 close on CIFAR-10.

Width-ratio cells are local off-budget checks around each iso width. The fixed-budget claim comes from the 1.0x column.

Depth by width ratio

CIFAR-100 accuracy

Iso-budget line

1.0x width · mean ±1 SD over 3 seeds

Interpretation

Extra width helps within each row, but the iso-budget column answers a different question. At the same target budget, the deeper plain networks were harder to train.

Dataset comparison

CIFAR-100 makes the depth penalty sharper.

CIFAR-10 shows a D=14 and D=20 plateau above D=8 and D=26. CIFAR-100 shows a clearer peak at D=14 and a larger deep-plain drop.

The table reports three-seed means from the plain CNN iso row. Accuracy and gap are percentages.

CIFAR-10 and CIFAR-100 iso rows

test accuracy
Dataset Depth Width Params Accuracy Gap

Gradient and trainability diagnostics

Deeper plain models converge more slowly.

Convergence epoch rises monotonically with depth at the iso budget. The gradient-ratio proxy does not collapse in a simple way, so the diagnosis is broader than one vanishing-gradient symptom.

Convergence is the epoch where online train accuracy reaches 99% of its own final value.

Convergence epoch by depth

plain CNN iso budget

Deep-side signal

D=8 mean 111.5
D=26 mean 129.0
Depths 4
Seeds per plain cell 3

Readout

The deepest plain endpoint has the slowest convergence and the weakest CIFAR-100 accuracy at the matched parameter budget.

Mechanism checks

The two extremes fail differently.

Residual connections improve the deepest model’s accuracy. The shallow model needs more parameters to reach similar accuracy.

ResNet controls are diagnostic interventions, not a fully powered architecture benchmark.

Plain versus ResNet

CIFAR-100 matched 150

Shallow-wide compensation

CIFAR-100

Mechanism notes

Residual recovery

D=26 CIFAR-100 moves from 73.66% plain to 78.00% with matched 150-epoch residual reruns.

Shallow cost

D=8 reaches 78.38% only at 19.96M parameters, roughly matching D=14 at 11.32M.

Fixed optimizer protocol

The gap varies while SGD noise scale is fixed.

All plain-grid runs share the same learning rate, batch size, momentum, data size, and scalar noise-scale value. Gap variation therefore cannot be reduced to the scalar g value alone.

Fixed value: g = eta N / B(1 - beta) = 390.625 for every plain-grid run.

Gap proxy by width ratio

CIFAR-100, fixed g

What the fixed protocol controls

The architectures share the same scalar noise-scale value. Accuracy gaps still vary, so that value alone does not explain them.

Paper figure set

Figures from the paper.

The interactive views above are generated from the same committed aggregate data used by the paper figures.

The report keeps only the six public figures used here.

Iso-parameter verification chart

Iso-parameter verification

Budget matching across the plain CNN grid.

CIFAR-100 accuracy heatmap

CIFAR-100 accuracy

The clearest D=14 peak in the frontier.

Trainability diagnostics

Trainability diagnostics

Convergence slows as plain depth rises.

ResNet control chart

ResNet control

Residual pathways recover the deep endpoint.

Noise scale check

Noise-scale check

Gap variation remains at fixed scalar g.

Gradient ratio heatmaps

Gradient proxy

The deep penalty is broader than one simple collapse.

Data and code

Saved experiment results.

The charts use aggregate CSV results committed to the repository.

Run `python scripts/build_report_data.py` after updating CSV summaries.