ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design

The paper observes that neural network architecture design is mostly guided by an indirect proxy for efficiency, computation complexity measured in FLOPs, even though the direct metric that users care about, such as speed, also depends on factors like memory access cost and platform characteristics. The authors therefore propose evaluating the direct metric on the target platform and run a series of controlled experiments to isolate what actually drives efficiency.

From these experiments they derive several practical guidelines for efficient network design and use them to build a new architecture, ShuffleNet V2. Comprehensive ablation studies verify that the resulting model is state-of-the-art in terms of the tradeoff between speed and accuracy, indicating that optimizing for direct, platform-aware metrics yields better real-world performance than optimizing FLOPs alone.