efficient net

EfficientNet is a family of convolutional neural networks that achieves state-of-the-art accuracy with significantly fewer parameters and FLOPs through compound scaling—simultaneously scaling network depth, width, and resolution in a principled manner. Key innovation: compound scaling method—instead of arbitrarily scaling one dimension (deeper, wider, or higher resolution), scale all three dimensions with fixed ratios determined by grid search. Scaling formula: depth d = α^φ, width w = β^φ, resolution r = γ^φ, where α, β, γ are constants (α·β²·γ² ≈ 2) and φ is compound coefficient. Architecture: EfficientNet-B0 (baseline—7.8M parameters, 0.39B FLOPs) designed via neural architecture search (NAS) using mobile inverted bottleneck (MBConv) blocks with squeeze-and-excitation. Family: B0 through B7 (scaling φ from 0 to 2.6)—B7 achieves 84.4% ImageNet top-1 with 66M parameters (vs. 145M for ResNet-152). MBConv blocks: (1) depthwise separable convolutions (reduce parameters), (2) inverted residuals (expand then compress), (3) SE attention (channel-wise recalibration). Advantages: (1) superior accuracy-efficiency trade-off (10× fewer parameters than previous SOTA), (2) scales well (consistent improvements from B0 to B7), (3) transfer learning (excellent pre-trained features). Applications: (1) mobile/edge deployment (B0-B2 for real-time inference), (2) cloud inference (B3-B5 for accuracy), (3) research (B6-B7 for benchmarks). Variants: EfficientNetV2 (faster training, better parameter efficiency), EfficientDet (object detection). EfficientNet demonstrated that principled scaling is more effective than ad-hoc architecture design, influencing subsequent efficient architecture research.