ShuffleNet Units are building blocks of the ShuffleNet architecture that use channel shuffle operations between grouped convolutions — solving the information isolation problem of grouped convolutions by rearranging channels so each group receives input from all previous groups.

How Do ShuffleNet Units Work?

• Grouped 1×1 Conv: Reduce channels via grouped pointwise convolution (efficient but isolates groups).
• Channel Shuffle: Reshape channels into $(G, C/G)$ -> transpose -> flatten. Now each group has channels from all original groups.
• Depthwise Conv: 3×3 depthwise convolution for spatial processing.
• Grouped 1×1 Conv: Another grouped pointwise to recover channel dimension.
• Paper: Zhang et al. (2018).

Why It Matters

• Cross-Group Communication: Channel shuffle bridges the information gap between groups without the cost of full 1×1 convolution.
• Extreme Efficiency: Designed for <100 MFLOPs models (smartwatch, IoT devices).
• v2: ShuffleNetV2 further optimizes for actual inference speed (not just FLOPs).

ShuffleNet Units are efficient blocks with channel shuffling — solving grouped convolution's isolation problem with a zero-computation rearrangement.