In this paper we describe a new mobile architecture, MobileNetV2, that improves the state of the art performance of mobile models on multiple tasks and benchmarks. Our main contribution is a novel layer module: the inverted residual with linear bottleneck. This module takes as input a low-dimensional compressed representation, first expands to high dimension, filters with a lightweight depthwise convolution, and then projects back to a low-dimensional representation with a linear convolution. Shortcut connections are between the thin bottleneck layers, contrasting with traditional residual networks that connect expanded representations.

Neural networks have revolutionized many areas of machine intelligence, but the computational demands of state-of-the-art models often exceed the capabilities of mobile and embedded devices. MobileNetV1 introduced depthwise separable convolutions as a computationally efficient alternative to standard convolutions. However, its architecture lacks residual connections and uses ReLU activations in low-dimensional spaces, potentially destroying useful information. MobileNetV2 addresses these issues with two key innovations: inverted residual structures where shortcut connections are between thin bottleneck layers, and the removal of non-linearities in narrow layers to preserve representational capacity.

Efficient neural network design has been pursued through several approaches: network pruning removes redundant connections, quantization reduces numerical precision, and knowledge distillation transfers knowledge from large to small networks. Architectural innovations include depthwise separable convolutions (MobileNetV1), ShuffleNet's channel shuffling, and SqueezeNet's fire modules. Residual connections (ResNet) and dense connections (DenseNet) improve gradient flow. MobileNetV2 uniquely combines inverted residuals with linear bottlenecks, a combination not previously explored.

The building block of MobileNetV2 is the inverted residual bottleneck. Unlike traditional residual blocks that have a "wide -> narrow -> wide" structure, the inverted residual follows a "narrow -> wide -> narrow" pattern. Specifically, a 1x1 convolution expands the input by a factor of t (expansion factor, typically t=6), then a 3x3 depthwise convolution filters the expanded representation, and finally a 1x1 linear convolution projects back to the low-dimensional bottleneck. Shortcut connections link the narrow input to the narrow output, keeping the residual connection in the low-dimensional space.

A critical theoretical insight motivates the linear bottleneck design. The authors establish that ReLU activations can destroy information when applied to low-dimensional manifolds. Specifically, "if the manifold of interest remains non-zero volume after ReLU transformation, it corresponds to a linear transformation" — meaning ReLU either destroys information (by zeroing out negative values) or acts as an identity (when all values are positive). In the narrow bottleneck where dimensionality is low, the probability of information loss through ReLU is high. Therefore, the final projection layer uses no activation function (linear), preserving all information in the compressed representation.

The complete MobileNetV2 architecture consists of an initial 32-filter standard convolution, followed by 19 inverted residual bottleneck layers, and concludes with a 1x1 convolution, global average pooling, and classification layer. The network uses ReLU6 activation (capped at 6) for robustness in low-precision (fixed-point) computation. A width multiplier parameter allows scaling the number of channels to trade accuracy for efficiency. The design also enables memory-efficient inference: since only the narrow bottleneck tensors need to be stored between layers, peak memory consumption is significantly reduced compared to architectures that store wide intermediate activations.

ImageNet classification: MobileNetV2 achieves 72.0% top-1 accuracy with 3.4M parameters and 300M multiply-adds, outperforming MobileNetV1 (70.6%) at comparable computational cost. The 1.4x width multiplier variant reaches 74.7% accuracy with 585M MAdds. Object detection on COCO: SSDLite with MobileNetV2 achieves 22.1% mAP with only 0.8B operations, described as "20x more efficient and 10x smaller" than YOLOv2 while maintaining competitive accuracy. Semantic segmentation on PASCAL VOC: Mobile DeepLabv3 achieves 75.32% mIOU using just 2.75B multiply-adds.

We have introduced MobileNetV2, a mobile architecture based on inverted residual structures with linear bottlenecks. The design is grounded in a theoretical understanding of how ReLU affects information flow in low-dimensional spaces. The resulting architecture achieves state-of-the-art efficiency across classification, detection, and segmentation tasks while remaining implementable with standard framework operations, requiring no special hardware or software support. MobileNetV2 provides a simple, theoretically grounded architecture for efficient deployment on mobile and embedded devices.