How ModelArts Powers AI Development and Seamless Edge‑Cloud Deployment

1. ModelArts Current Status and Development Features

ModelArts is Huawei's full‑stack, all‑scenario AI solution portal for users and developers, offering massive data preprocessing, semi‑automatic labeling, large‑scale distributed training, automated model generation, and on‑demand edge‑cloud deployment to accelerate model creation and lifecycle management.

1.2 Data Processing

The platform provides data collection, filtering, labeling, and dataset version management, supporting automated and semi‑automated workflows that reduce the time spent on data preparation and annotation.

1.2.1 Algorithm Development

ModelArts includes MoXing, a lightweight distributed framework built on TensorFlow, PyTorch, and MXNet, which simplifies hyperparameter tuning, integrates multiple tuning strategies, and enables automatic conversion of single‑node code to distributed execution with high performance.

1.2.2 Model Training Optimizations

Key optimizations include automatic mixed‑precision training, dynamic hyper‑parameter adjustment (batch size, image size, momentum), automatic gradient fusion, BP‑bubble adaptive compute‑communication scheduling, high‑performance communication libraries (nstack, HCCL), data‑model hybrid parallelism, and multi‑level caching.

1.2.3 Model Deployment

ModelArts supports edge‑cloud collaborative deployment, allowing models to run on cloud, edge, and device scenarios. It offers model splitting, task partitioning, version management, model search, reuse, and security controls such as permission authentication and billing.

1.2.4 AI Marketplace

The AI Marketplace built on ModelArts provides a community for sharing models, APIs, datasets, and competitions, fostering a secure and open environment for researchers, developers, and enterprises to accelerate AI product development.

1.2.5 Auto‑Learning

Through automated machine learning, ModelArts enables users without algorithm expertise to generate models via transfer learning and neural architecture search, supporting image classification, object detection, predictive analysis, and sound classification without writing code.

2. Edge‑Cloud Collaborative Deployment for Invisible Recognition

The platform underpins invisible payment technology, which binds unique physical identifiers to bank accounts, enabling authentication without IDs or passwords and achieving sub‑2‑second transaction speeds. ModelArts integrates optical, acoustic, and wave‑based recognition on Ascend/MindSpore edge devices, with edge processing, edge services, and cloud orchestration ensuring high precision and low latency.

A practical case is the "Laborer Harbor" intelligent waste‑sorting system deployed at Huawei Cloud Shanghai 2019 HC conference. An Android device captures waste images, sends them via Wi‑Fi/5G to ModelArts for inference, and returns classification results based on Shenzhen waste‑sorting standards, demonstrating scalable edge‑cloud AI deployment.

2.1 Technical Details

The solution uses TensorFlow 1.6 with models such as SENet, ResNet‑50/152, and SEResNet. After baseline tuning, ResNet‑50 achieved 0.8756 accuracy on the validation set, later improved to >0.95 with cosine‑annealing learning rate schedules.

3. Ecosystem Outlook for Invisible Recognition

Invisible payment serves as an entry point to a broader AI ecosystem, generating high‑quality transaction data that enriches user profiling and enables new services for consumers, governments, and enterprises. Example applications include unmanned parking fee robots using 3‑D vehicle recognition and next‑generation ETC systems that replace traditional RFID tags with AI‑driven video recognition, improving throughput and reducing infrastructure costs.

4. Conclusion

ModelArts represents one of the fastest AI platforms in terms of compute, inference, and accuracy. Its continuous evolution and edge‑cloud capabilities are poised to advance invisible payment and related AI applications, delivering more convenient and intelligent services for society.