Compound model scaling¶
A neural-network family design rule that coordinates changes in multiple architecture dimensions as the resource budget changes.
Core Idea¶
Compound model scaling makes a family of related neural networks by changing at least two architecture dimensions together as the resource budget changes. A declared rule says how the changes fit together. EfficientNet coordinates depth, channel width and image resolution in image classifiers; the large Scaled-YOLOv4 detector family first links input size with network stages, then adjusts depth and width for its deployment conditions. The shared idea is the coordinated architecture rule, not one formula or a guarantee of better performance.[ref-8dc917e3d83c][ref-abf2238fa9c7]
Scope of Application¶
Use the term for neural-network architecture families with a baseline, two or more coupled dimensions, a resource or deployment condition, and related variants generated by the rule. Depth, width, resolution and stage count are examples, not a mandatory list. A new model family must define its own coupling and test its results. The method does not automatically cover every machine-learning model or every pair of quantities that grows together.[ref-8dc917e3d83c][ref-abf2238fa9c7]
Clarity¶
Ask: What is the baseline? Which architecture dimensions move together? What rule couples them? Which resource condition selects a variant? The resulting variants instantiate the method; accuracy and speed measurements evaluate them separately. Changing just one dimension is single-axis scaling, and changing several knobs without a family rule is insufficient. EfficientNet's coefficient and approximate FLOP-doubling condition specify its particular classifier progression; they are not requirements for Scaled-YOLOv4.[ref-8dc917e3d83c][ref-abf2238fa9c7]
Manages Complexity¶
A rule narrows a large design space into a progression whose variants can be compared. If performance disappoints, one can inspect whether a limiting dimension was left behind or whether the coupling spends too much of the latency budget. The rule makes the choice legible, but does not prove it optimal.[ref-8dc917e3d83c][ref-abf2238fa9c7]
Abstract Reasoning¶
Hold the baseline and task fixed. Trace at least two changed architecture dimensions across resource levels, and check that their relation follows a declared rule. Then evaluate outcomes under comparable data and compute or latency conditions. EfficientNet reports a controlled comparison with single-axis changes in its own tests. The Scaled-YOLOv4 paper supplies a separate detector rationale and measured results, but not that same matched comparison for its P5–P7 family.[ref-8dc917e3d83c][ref-abf2238fa9c7]
Knowledge Transfer¶
Within neural architecture design, transfer the question of which dimensions must be coordinated under a budget. Do not transfer EfficientNet's coefficients, grid search or measured benefit as universal instructions.[ref-8dc917e3d83c][ref-abf2238fa9c7] In another domain, the general input-rule-output pattern belongs to the parent Prime Transformation; this named method remains tied to neural architectures.
Example¶
EfficientNet starts from B0 and advances related B1–B7 classifiers by coordinating depth, width and resolution with a chosen compound coefficient. For an unlike detector case, Scaled-YOLOv4 expands P5 to P6 and P7 with an input-size/stage relation and later depth and width adjustments under inference-time constraints. In both, the baseline, coupled dimensions, resource condition and resulting family are identifiable; reported accuracy and speed are evaluation, not defining roles.[ref-8dc917e3d83c][ref-abf2238fa9c7]
Relationships to Other Abstractions¶
Current abstraction Compound model scaling Domain-specific
Parents (1) — more general patterns this builds on
-
Compound model scaling is a kind of Transformation Prime
Compound model scaling is a Transformation specialized to coordinated restructuring of a neural architecture family under a resource condition.
Hierarchy path (1) — routes to 1 parentless root
- Compound model scaling → Transformation → Function (Mapping)
Neighborhood in Abstraction Space¶
Compound model scaling sits in a sparse region of the domain-specific corpus (98th percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.
Family — Unclustered & Miscellaneous (2721 abstractions)
Nearest neighbors
- Benford's Law — 0.75
- Ridgeline Plot — 0.75
- Bateman–Mukai Method — 0.75
- Condition Number — 0.75
- Tag SNP — 0.75
Computed from structural-signature embeddings · 2026-10-17
Not to Be Confused With¶
A neural scaling law describes an observed relation between resources and performance; it does not specify the architecture changes that generate variants. Neural architecture search is broader than one declared family progression. A valid compound rule also does not certify efficiency: another architecture might outperform its variants.[ref-8dc917e3d83c][ref-abf2238fa9c7] Transformation is the strict graph parent because each instance applies a rule to a baseline architecture and preserves family lineage; Transformation also covers non-neural and single-axis changes.
References¶
[^ref-8dc917e3d83c]: Mingxing Tan and Quoc V. Le, “EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks,” Proceedings of the 36th International Conference on Machine Learning (ICML 2019), pp. 6105–6114. Proceedings record: https://proceedings.mlr.press/v97/tan19a.html ; full original author version v5 inspected at https://arxiv.org/html/1905.11946 , DOI: 10.48550/arXiv.1905.11946, especially §§3.1–3.3, 4, 5.1 and 6. [^ref-abf2238fa9c7]: Chien-Yao Wang, Alexey Bochkovskiy and Hong-Yuan Mark Liao, “Scaled-YOLOv4: Scaling Cross Stage Partial Network,” Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR 2021), pp. 13029–13038. Proceedings record: https://openaccess.thecvf.com/content/CVPR2021/html/Wang_Scaled-YOLOv4_Scaling_Cross_Stage_Partial_Network_CVPR_2021_paper.html ; full original inspected at https://arxiv.org/pdf/2011.08036 , DOI: 10.1109/CVPR46437.2021.01283, especially §§2.2, 3.3, 4.2–4.3 and 5.