> ML_LITERATURE // HOWARD-2017-MOBILENETS-EFFICIENT-CONVOLUTIONAL-NEURAL-NETWORKS-MOBILE-VISION_v1.0
MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications
Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, Hartwig Adam · arXiv preprint (2017)
seminal-architecture2017industry-standardthirdPartyReproduced
Principal Contribution
Factorized standard convolutions into depthwise and pointwise convolutions, reducing compute and parameters by 8x-9x with minimal accuracy loss.
Operational Relevance
The industry-standard convolutional backbone deployed on billions of iOS and Android smartphones for edge real-time vision.
Assumptions
- Spatial filtering and cross-channel feature combination can be completely decoupled without sacrificing representational capacity
Limitations
- Depthwise convolutions have low arithmetic intensity, requiring memory-bandwidth-optimized mobile runtime runtimes (TFLite, CoreML)
Connected Algorithms, Architectures & Tools
Related Algorithms:
Related Architectures:
Implementing Libraries: