MobileNet

What MobileNet Does

MobileNet is a family of compact convolutional networks designed for phones, embedded systems, and other latency- or power-constrained devices. It can serve as a classifier or as the backbone of detection and segmentation models.

MobileNetV1 was introduced in MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications. V2 added inverted residuals and linear bottlenecks; V3 combined hardware-aware architecture search, squeeze-and-excitation, and hard-swish.

Core Idea

A standard convolution learns spatial patterns and mixes channels in one expensive operation. MobileNetV1 factors it into a depthwise convolution that filters each channel separately and a pointwise 1×11\times1 convolution that mixes channels.

How It Works

1. Separate Spatial and Channel Processing

For a D×DD\times D kernel, CinC_{in} input channels, CoutC_{out} output channels, and an H×WH\times W feature map:

  • Standard convolution: D2CinCoutHWD^2 C_{in} C_{out} HW
  • Depthwise convolution: D2CinHWD^2 C_{in} HW
  • Pointwise convolution: CinCoutHWC_{in} C_{out} HW

The separable-to-standard cost ratio is 1/Cout+1/D21/C_{out}+1/D^2. With a 3×33\times3 kernel and many output channels, this approaches a ninefold arithmetic reduction, though measured latency depends on hardware and kernels.

2. Choose an Accuracy–Cost Point

V1's width multiplier α\alpha scales channel counts; its resolution multiplier ρ\rho scales spatial dimensions. Compute therefore changes approximately with α2ρ2\alpha^2\rho^2, subject to channel rounding and fixed layers.

3. Preserve Information in MobileNetV2

An inverted residual block expands a narrow input, applies a depthwise convolution in the wide space, then linearly projects back to a narrow output. A residual connection is used when shape and stride permit. Keeping the projection linear avoids destroying information in the low-dimensional bottleneck.

4. Optimize for Real Devices in MobileNetV3

MobileNetV3 uses platform-aware neural architecture search, squeeze-and-excitation for channel recalibration, and hard-swish:

h-swish(x) = x · ReLU6(x + 3) / 6

Representative ImageNet configurations range from MobileNetV3-Small (about 2.5M parameters) to MobileNetV3-Large (about 5.4M); exact accuracy and latency depend on training recipe and device.

Why It Works

  • Spatial filtering does not always need to be repeated independently for every output channel.
  • Cheap 1×11\times1 convolutions handle most channel interaction.
  • Width and resolution controls let deployment teams target a measured resource budget.
  • V2 and V3 place capacity where it offers useful accuracy without restoring the full cost of standard convolutions.

Limitations

  • Fewer operations do not guarantee lower latency: depthwise kernels can be memory-bound or poorly optimized.
  • Small variants trade away capacity and may struggle on fine-grained classes or dense prediction.
  • Width, resolution, quantization, and runtime must be benchmarked on the target device.
  • Training small models from scratch can be sensitive; pretrained initialization and distillation are often helpful.

Evolution

Related

Sources