We present a study on rethinking the bottleneck structure for efficient mobile network design. We revisit the widely used inverted residual bottleneck in MobileNetV2 and show that the expansion-depthwise-projection design may not be optimal. We propose a new sandglass block that reverses the structure: narrow-wide-narrow becomes wide-narrow-wide, performing identity mapping on high-dimensional features. This design outperforms MobileNetV2 on ImageNet with fewer FLOPs.

The inverted residual block in MobileNetV2 uses an expansion layer to increase channel dimensions, a depthwise separable convolution, and a projection layer to reduce dimensions. The shortcut connection operates on the low-dimensional bottleneck, which we argue discards important information and limits gradient flow. We propose to perform the shortcut on the high-dimensional representation instead, maintaining richer information across layers.

From an information bottleneck perspective, the inverted residual's shortcut on low-dimensional features forces the network to compress all information into a narrow channel. This is particularly problematic when spatial features are important, as the projection step inevitably loses spatial details. Our sandglass design avoids this by maintaining the high-dimensional representation in the shortcut path, allowing spatial information to flow freely across blocks.

The proposed sandglass block reverses the inverted residual structure. It starts with high-dimensional input features, applies depthwise convolutions to capture spatial information, then reduces dimensions with a linear bottleneck, and finally expands back to the original high dimension. The residual connection bridges the high-dimensional input and output, preserving maximum information through the shortcut. This design maintains the efficiency of depthwise separable convolutions while improving information flow.

We provide both theoretical analysis and empirical evidence for the sandglass design. From an information theory perspective, the shortcut on high-dimensional features preserves more mutual information between input and output. We also show that gradient magnitude through the sandglass block is larger than through the inverted residual, indicating better gradient flow during training.

On ImageNet classification, our MobileNeXt with sandglass blocks achieves 74.02% top-1 accuracy with 300M FLOPs, compared to MobileNetV2's 72.0% at 300M FLOPs — a 2% improvement at the same computational cost. On COCO object detection, MobileNeXt as backbone improves SSDLite mAP by 1.7% over MobileNetV2. The model also shows better transferability to downstream tasks.

Ablation studies further validate the design choices. Comparing shortcut on high-dimensional vs low-dimensional features shows a consistent +1.5% advantage for high-dimensional shortcuts across different model sizes. The depthwise convolution placement (before vs after bottleneck) also matters: placing it before the reduction step captures richer spatial information, contributing +0.6% accuracy.

We have challenged the inverted residual bottleneck design and shown that reversing the structure — performing shortcuts on high-dimensional features — yields better mobile networks. The sandglass block is a simple yet effective alternative that improves both accuracy and efficiency. We hope this work encourages the community to revisit established architectural designs with fresh perspectives.