GhostNet: More Features from Cheap Operations (1911.11907v2)

Abstract: Deploying convolutional neural networks (CNNs) on embedded devices is difficult due to the limited memory and computation resources. The redundancy in feature maps is an important characteristic of those successful CNNs, but has rarely been investigated in neural architecture design. This paper proposes a novel Ghost module to generate more feature maps from cheap operations. Based on a set of intrinsic feature maps, we apply a series of linear transformations with cheap cost to generate many ghost feature maps that could fully reveal information underlying intrinsic features. The proposed Ghost module can be taken as a plug-and-play component to upgrade existing convolutional neural networks. Ghost bottlenecks are designed to stack Ghost modules, and then the lightweight GhostNet can be easily established. Experiments conducted on benchmarks demonstrate that the proposed Ghost module is an impressive alternative of convolution layers in baseline models, and our GhostNet can achieve higher recognition performance (e.g. $75.7\%$ top-1 accuracy) than MobileNetV3 with similar computational cost on the ImageNet ILSVRC-2012 classification dataset. Code is available at https://github.com/huawei-noah/ghostnet

GhostNet: More Features from Cheap Operations

The paper "GhostNet: More Features from Cheap Operations" presents a new approach to addressing the computational limitations inherent in deploying convolutional neural networks (CNNs) on embedded devices. The primary innovation introduced is the Ghost module, a technique designed to generate multiple feature maps through computationally inexpensive linear operations, effectively addressing the redundancy observed in feature maps of deep networks.

Key Contributions

1. Ghost Module: The core contribution is the Ghost module, which reduces the conventional reliance on a large number of convolutional filters. Feature maps are generated in two stages: the first using standard convolutions to produce intrinsic feature maps, and the second applying simple linear transformations to produce additional, so-called ghost feature maps. This approach significantly reduces the number of parameters and FLOPs necessary for the CNN while maintaining a high level of performance.
2. Ghost Bottlenecks and GhostNet: Building on the Ghost module, the authors propose Ghost bottlenecks, which consist of stacked Ghost modules. These bottlenecks serve as the foundation for GhostNet, a lightweight neural architecture intended for efficient deployment on mobile devices. GhostNet is showcased in multiple variants to trade off between performance and computational efficiency, evaluated primarily through image classification tasks.

Experiments and Results

Experimental Validation: The effectiveness of GhostNet and its components is validated across several tasks and datasets, including CIFAR-10, ImageNet ILSVRC-2012, and MS COCO.

• CIFAR-10: By embedding Ghost modules in standard architectures such as VGG-16 and ResNet-56, the researchers demonstrate significant reductions in computational cost without degradation in accuracy. For instance, Ghost-VGG-16 achieves 93.7% accuracy with a 2× reduction in FLOPs.
• ImageNet: GhostNet is compared against multiple state-of-the-art efficient networks. For example, GhostNet achieves 75.7% top-1 accuracy on the ImageNet dataset while requiring fewer FLOPs than MobileNetV3, illustrating its superior efficiency and performance balance.
• Object Detection: The generality of GhostNet is demonstrated through object detection tasks on the MS COCO dataset, where GhostNet-based backbones perform comparably to MobileNetV2 and MobileNetV3 but with significantly lower computational costs.

Implications and Future Directions

The practical implications of GhostNet are substantial for real-world applications requiring efficient computation on resource-constrained devices. Embedded devices, such as smartphones and autonomous vehicles, benefit significantly from reduced computational requirements, enabling more complex and accurate models to run in environments with limited computational power.

The theoretical contribution of identifying and exploiting feature map redundancy introduces a new perspective in CNN design. Rather than avoiding redundancy, GhostNet incorporates it as an integral component, thereby optimizing the computational process through more efficient operations.

Future developments in AI could further enhance this methodology through several avenues:

• Diverse Linear Operations: Investigating a broader array of linear operations beyond current depthwise convolutions could yield even more efficient methodologies for generating ghost feature maps.
• Architecture Search: Automated neural architecture search tailored to GhostNet principles might uncover even more optimal structures for diverse applications.
• Hardware-Specific Optimizations: Fine-tuning Ghost modules to leverage specific hardware accelerations available in modern CPUs and GPUs could enhance real-world deployment efficiency even further.

Conclusion

This paper "GhostNet: More Features from Cheap Operations" contributes significantly to the field of efficient neural network design. The introduction of the Ghost module, Ghost bottlenecks, and the overall GhostNet architecture provides a practical, efficient alternative for developing lightweight CNNs while maintaining a balance between computational efficiency and high recognition performance. The demonstrated success across various benchmarks places GhostNet as a competitive contender for deployment in computationally constrained environments.