EfficientNet Scaling

Unlike traditional approaches that scale a single dimension, EfficientNet uses a compound coefficient φ to uniformly scale all three dimensions:

• Depth (d): Number of layers, scaled by α^φ
• Width (w): Number of channels per layer, scaled by β^φ
• Resolution (r): Input image size, scaled by γ^φ

Where α, β, γ are constants determined by a small grid search on the baseline model, satisfying the constraint α · β² · γ² ≈ 2. This ensures total FLOPS increase by approximately 2^φ for each step.

The foundation of the family is EfficientNet-B0, a mobile-sized architecture discovered through neural architecture search (NAS) optimized for both accuracy and FLOPS:

• Built primarily with mobile inverted bottleneck MBConv blocks
• Incorporates squeeze-and-excitation optimization for channel-wise attention
• Achieves 76.3% top-1 accuracy on ImageNet with only 5.3M parameters
• Serves as the seed architecture for all subsequent scaling

The baseline itself represents a significant improvement over hand-crafted mobile architectures like MobileNetV2.

Applying compound scaling produces a family of models from B0 to B7 that dominate the accuracy-efficiency Pareto frontier:

• EfficientNet-B7 achieves 84.3% top-1 accuracy on ImageNet with 66M parameters
• Outperforms previous state-of-the-art GPipe while using 8.4x fewer parameters
• Consistently achieves higher accuracy than comparable models at every FLOPS budget
• The scaling method transfers effectively to other tasks including object detection and semantic segmentation

The family demonstrates that balanced scaling is superior to arbitrary single-dimension increases.

The optimal scaling coefficients are determined through a two-step process:

Step 1: Fix φ=1 and perform a small grid search over α, β, γ on the baseline model to find values satisfying the FLOPS constraint α · β² · γ² ≈ 2

Step 2: Fix α, β, γ and scale up φ to produce larger variants

This decoupling ensures the scaling rule is discovered once on a small model and then extrapolated to larger architectures without expensive per-model tuning. The search space is small enough to be practical while capturing the critical interactions between dimensions.

EfficientNet models demonstrate exceptional transfer learning capabilities across diverse computer vision tasks:

• Achieved state-of-the-art on CIFAR-100 (91.7%) and Flowers (98.8%) with fine-tuning
• Scaled EfficientNet-B7 attains 55.1% mAP on COCO object detection with fewer parameters than ResNet-based detectors
• The compound-scaled representations generalize well because balanced dimensions capture multi-scale features effectively
• Particularly beneficial for edge deployment scenarios where both accuracy and computational efficiency are critical

The scaling methodology produces features that are inherently robust across domains.

The compound coefficient formalizes an intuitive observation about ConvNet scaling:

• Higher resolution images benefit from deeper networks with larger receptive fields to capture fine-grained patterns
• Wider networks provide more channels to represent the increased detail from higher resolution inputs
• Scaling only one dimension yields diminishing returns — a wider but shallow network cannot effectively process high-resolution inputs
• The balanced approach ensures all dimensions grow proportionally, maximizing the return on each additional FLOPS

This principle applies broadly beyond EfficientNet and informs modern architecture design for both vision and signal processing tasks.