EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks .

Abstract

EfficientNet systematically studies model scaling and proposes compound scaling: uniformly scaling depth, width, and resolution.

Key Insight

Scaling dimensions independently is suboptimal. Compound scaling:

• Depth: Number of layers
• Width: Number of channels
• Resolution: Input image size

Use neural architecture search to find the baseline, then scale all three dimensions with a fixed ratio.

Impact

EfficientNet-B7 achieved state-of-the-art ImageNet accuracy with 8.4x fewer parameters than the previous best model. Sets new standard for efficiency.