MobileNet is a family ofconvolutional neural network (CNN) architectures designed forimage classification,object detection, and other computer vision tasks. They are designed for small size, low latency, and low power consumption, making them suitable for on-device inference andedge computing on resource-constrained devices likemobile phones andembedded systems. They were originally designed to be run efficiently on mobile devices withTensorFlow Lite.

The need for efficient deep learning models on mobile devices led researchers atGoogle to develop MobileNet. As of October 2024^[update], the family has four versions, each improving upon the previous one in terms of performance and efficiency.

MobileNetV1 was published in April 2017.^[1][2] Its main architectural innovation was incorporation ofdepthwise separable convolutions. It was first developed by Laurent Sifre during an internship atGoogle Brain in 2013 as an architectural variation onAlexNet to improve convergence speed and model size.^[3]

The depthwise separable convolution decomposes a single standard convolution into two convolutions: a depthwise convolution that filters each input channel independently and a pointwise convolution (${\displaystyle 1\times 1}$ convolution) that combines the outputs of the depthwise convolution. This factorization significantly reduces computational cost.

The MobileNetV1 has two hyperparameters: awidth multiplier${\displaystyle \alpha }$ that controls the number of channels in each layer. Smaller values of${\displaystyle \alpha }$ lead to smaller and faster models, but at the cost of reduced accuracy, and aresolution multiplier${\displaystyle \rho }$, which controls the input resolution of the images. Lower resolutions result in faster processing but potentially lower accuracy.

MobileNetV2 was published in March 2019.^[4][5] It usesinverted residual layers andlinear bottlenecks.

Inverted residuals modify the traditional residual block structure. Instead of compressing the input channels before the depthwise convolution, theyexpand them. This expansion is followed by a${\displaystyle 1\times 1}$ depthwise convolution and then a${\displaystyle 1\times 1}$ projection layer that reduces the number of channels back down. This inverted structure helps to maintain representational capacity by allowing the depthwise convolution to operate on a higher-dimensional feature space, thus preserving more information flow during the convolutional process.

Linear bottlenecks removes the typical ReLU activation function in the projection layers. This was rationalized by arguing that that nonlinear activation loses information in lower-dimensional spaces, which is problematic when the number of channels is already small.

MobileNetV3 was published in 2019.^[6][7] The publication included MobileNetV3-Small, MobileNetV3-Large, and MobileNetEdgeTPU (optimized forPixel 4). They were found by a form ofneural architecture search (NAS) that takes mobile latency into account, to achieve good trade-off between accuracy and latency.^[8][9] It used piecewise-linear approximations ofswish andsigmoidactivation functions (which they called "h-swish" and "h-sigmoid"),squeeze-and-excitation modules,^[10] and the inverted bottlenecks of MobileNetV2.

MobileNetV4 was published in September 2024.^[11][12] The publication included a large number of architectures found by NAS. Compared to the architectural modules used in V3, the V4 series included the "universal inverted bottleneck", which includes both inverted residual and inverted bottleneck as special cases, and attention modules with multi-query attention.^[13]

• Convolutional neural network
• Deep learning
• TensorFlow Lite

• "models/research/slim/nets/mobilenet at master · tensorflow/models".GitHub. Retrieved2024-10-18.
• "Keras documentation: MobileNet, MobileNetV2, and MobileNetV3".Keras. RetrievedOctober 18, 2024.

• Computer vision
• Machine learning
• Google software

• CS1 errors: missing periodical
• Articles with short description
• Short description is different from Wikidata
• Articles containing potentially dated statements from October 2024
• All articles containing potentially dated statements