Vision · 2019 · scaling laws
EfficientNet
A family of vision networks derived from a compound scaling rule that balances depth, width and input resolution under a fixed compute budget.
ModLensVision
Interactive Diagram
Focus the lens
Click any component to read what it does. Signal direction follows the edges.
inputScaled Input. Resolution is a first-class scaling dimension.
Core idea
Scaling is a design variable. Instead of growing one dimension, grow all three with fixed ratios found by a small search — accuracy per FLOP becomes the objective.
Why it exists
Ad-hoc scaling wastes compute; a principled ratio extracts more accuracy from the same budget.
Mathematics
d=\alpha^\phi,\; w=\beta^\phi,\; r=\gamma^\phi,\; \alpha\beta^2\gamma^2 \approx 2
Data Flow
What moves through the system
- 01Input at a scaled resolution.
- 02MBConv blocks (depthwise separable convolutions) with squeeze-excitation.
- 03Compound-scaled stages balance depth, width, resolution.
- 04Pooling head for the task.
Strengths
- + Excellent accuracy/FLOP trade-off
- + Simple recipe to resize for deployment targets
Limitations
- − Scaling ratios are tuned to the base architecture
- − At its best in the small-to-mid compute regime where efficiency decides
Applications
- · Mobile and edge vision
- · Cost-sensitive production classifiers