mobilenetv2

**MobileNetV2** is the **second generation of MobileNet that introduces inverted residual blocks with linear bottlenecks** — expanding channels before depthwise convolution (inverted from ResNet's bottleneck) and removing the activation function after the final projection. **What Is MobileNetV2?** - **Inverted Residual**: Expand (1×1) -> Depthwise (3×3) -> Project (1×1). Expansion ratio $t = 6$. - **Linear Bottleneck**: No activation (ReLU) after the final 1×1 projection to prevent information loss in low-dimensional features. - **Skip Connection**: Residual connection between the narrow bottleneck features (not the expanded features). - **Paper**: Sandler et al. (2018). **Why It Matters** - **SSD/Detection**: The default mobile backbone for object detection and segmentation on device. - **Information Manifold**: The linear bottleneck insight — ReLU in low dimensions destroys information — is theoretically motivated. - **Industry Standard**: Used in TensorFlow Lite, MediaPipe, and countless mobile applications. **MobileNetV2** is **the inverted bottleneck revolution** — proving that expanding before filtering and projecting linearly produces better mobile features.