Efficient & searched architectures
Mobile-friendly building blocks, neural architecture search and principled model scaling.
| Concept | Introduced by | Years | Papers |
|---|---|---|---|
| Depthwise-separable convolutions Factor a convolution into per-channel and 1×1 steps to cut computation drastically. | MobileNets | 2017 | 0 |
| Neural architecture search Search for a network's building block automatically instead of designing it by hand. | NASNet | 2017–2019 | 2 |
| Inverted residuals & linear bottlenecks Expand, filter with depthwise conv, then project back down to thin residual bottlenecks. | MobileNetV2 | 2018 | 1 |
| Latency-aware architecture search Put measured on-device latency into the search objective. | MnasNet | 2018 | 0 |
| Compound model scaling Scale depth, width and resolution together with one coefficient. | EfficientNet | 2019–2021 | 5 |