PaperReading

MnasNet: Platform-Aware Neural Architecture Search for Mobile

May 2019

tl;dr: Search the neighborhood of MobileNetV2.

Overall impression

One of the main challenge of NAS is its vast search space. This paper uses MobilenetsV2 as a starting point and significantly reduces the search space. M stands for mobile.

The algorithm can be seen as an evolution algorithm, just a glorified for loop.

The performance is overtaken by FBNet also published at CVPR 2019, which uses differentiable optimization method instead of training a controller.

Key ideas

• Combine model Accuracy and latency into a cost function $ACC(model) \times [LAT(model)/TAR]^w$, with w ranging from -0.07 (in MobilenetsV3 Large, or MNasNet-A1) to -0.15 (in MobilenetsV3 Small). -0.07 was obtained by observing that empirically model accuracy improves by 5% when doubling latency.

Technical details

• Summary of technical details

Notes

• Questions and notes on how to improve/revise the current work