Searching for MobileNetV3 (1905.02244v5)

Abstract: We present the next generation of MobileNets based on a combination of complementary search techniques as well as a novel architecture design. MobileNetV3 is tuned to mobile phone CPUs through a combination of hardware-aware network architecture search (NAS) complemented by the NetAdapt algorithm and then subsequently improved through novel architecture advances. This paper starts the exploration of how automated search algorithms and network design can work together to harness complementary approaches improving the overall state of the art. Through this process we create two new MobileNet models for release: MobileNetV3-Large and MobileNetV3-Small which are targeted for high and low resource use cases. These models are then adapted and applied to the tasks of object detection and semantic segmentation. For the task of semantic segmentation (or any dense pixel prediction), we propose a new efficient segmentation decoder Lite Reduced Atrous Spatial Pyramid Pooling (LR-ASPP). We achieve new state of the art results for mobile classification, detection and segmentation. MobileNetV3-Large is 3.2\% more accurate on ImageNet classification while reducing latency by 15\% compared to MobileNetV2. MobileNetV3-Small is 4.6\% more accurate while reducing latency by 5\% compared to MobileNetV2. MobileNetV3-Large detection is 25\% faster at roughly the same accuracy as MobileNetV2 on COCO detection. MobileNetV3-Large LR-ASPP is 30\% faster than MobileNetV2 R-ASPP at similar accuracy for Cityscapes segmentation.

• The paper presents the integration of hardware-aware NAS and NetAdapt to optimize neural network architectures for mobile devices.
• The paper reports significant results, including a 3.2% accuracy improvement and a 20% latency reduction on MobileNetV3-Large compared to previous models.
• The paper demonstrates MobileNetV3's versatility by enhancing performance across image classification, object detection, and semantic segmentation tasks.

A Technical Review of "Searching for MobileNetV3"

The paper "Searching for MobileNetV3" presents substantial advancements in the development of efficient neural networks optimized for mobile devices. Authored by a team of researchers from Google AI and Google Brain, the paper introduces MobileNetV3, a series of models specifically tailored for high and low-resource use cases, thereby furthering the state of the art in mobile neural network performance.

Summary of Methodology

MobileNetV3's architecture is derived from a combination of complementary neural architecture search techniques and novel architecture design advancements. The paper details the process of optimizing these models using two main methodologies:

1. Hardware-Aware Network Architecture Search (NAS): This technique is employed to discover effective global network structures on a predefined search space. The search process accommodates different hardware constraints to obtain a balance between accuracy and latency.
2. NetAdapt Algorithm: This approach fine-tunes the layer configurations identified by NAS. NetAdapt modifies the number of filters in each layer to achieve precise balance between computational load and performance.

Additionally, the paper discusses various improvements in neural architecture that enhance the efficiency and performance of MobileNetV3. These enhancements include new forms of nonlinearity like hard-swish, integration of squeeze-and-excitation modules, and optimized handling of initial and final layers of the network.

Numerical Results

The comprehensive experiments presented in the paper highlight the efficacy of MobileNetV3 models across three key computer vision tasks: image classification, object detection, and semantic segmentation.

1. Image Classification:
   • MobileNetV3-Large improves the top-1 accuracy by 3.2% on the ImageNet dataset while reducing latency by 20% compared to MobileNetV2.
   • MobileNetV3-Small achieves 6.6% higher accuracy compared to a MobileNetV2 model with comparable latency.
2. Object Detection:
   • Utilizing MobileNetV3 as a backbone for SSDLite, the model delivers over 25% faster latency at a similar accuracy level compared to MobileNetV2 on the COCO dataset.
3. Semantic Segmentation:
   • With the new segmentation decoder Lite Reduced Atrous Spatial Pyramid Pooling (LR-ASPP), MobileNetV3-Large achieves 34% faster performance than MobileNetV2 while maintaining similar accuracy on the Cityscapes dataset.

Implications and Future Directions

The paper’s results hold significant implications for both theoretical and practical aspects of deep learning and edge computing:

• Efficient Model Design: MobileNetV3 demonstrates that careful blending of NAS techniques and manual architectural improvements can yield highly efficient models. This methodology can be further applied to other domains demanding low-latency AI applications like augmented reality and autonomous driving.
• Generalization Across Tasks: The successful adaptation of MobileNetV3 for various computer vision tasks suggests that the approach is not domain-specific but exhibits robust generalization capabilities. Future research might extend these techniques further to other forms of machine learning tasks, including natural language processing and reinforcement learning.
• Hardware-Aware Optimization: Continued exploration into hardware-aware NAS and automated fine-tuning models will likely provide even better optimization and adaptability to evolving hardware. This is particularly pertinent as new mobile processors with diverse computational capabilities and specialized AI accelerators emerge.

Conclusion

"Searching for MobileNetV3" rigorously details the creation of next-generation mobile-optimized neural networks, showcasing innovative methods to push the limits of efficiency and accuracy. By leveraging hybrid search techniques and strategic architectural optimizations, MobileNetV3 sets new benchmarks in mobile classification, detection, and segmentation tasks. The approaches and improvements discussed in the paper pave the way for future advancements in the field, encouraging further innovation in efficient deep learning model design.