efficientnet

**EfficientNet** is a **family of CNN architectures that uses a principled compound scaling method to uniformly scale network depth, width, and resolution** — achieving state-of-the-art accuracy at each efficiency level from mobile to server-scale. **What Is EfficientNet?** - **Baseline**: EfficientNet-B0 found by NAS (MnasNet-like search). - **Compound Scaling**: Jointly scale depth ($d = alpha^phi$), width ($w = eta^phi$), and resolution ($r = gamma^phi$) where $alpha cdot eta^2 cdot gamma^2 approx 2$. - **Family**: B0 through B7 (scaling factor $phi$ from 0 to 6). - **Paper**: Tan & Le (2019). **Why It Matters** - **Principled Scaling**: First to show that balanced scaling of all three dimensions outperforms scaling any one alone. - **Efficiency**: EfficientNet-B3 matches ResNet-152 accuracy with 8× fewer FLOPs. - **Standard**: Became the default CNN backbone for many vision tasks (2019-2021). **EfficientNet** is **the science of neural network scaling** — proving that balanced growth in depth, width, and resolution is the key to efficient accuracy.