Creating the AI world together with developers
At HUAWEI CONNECT 2018, Zheng Yelai, Huawei Vice President and President of Cloud BU, launched several platforms and tools for E2E AI development, including ModelArts, a faster inclusive AI development platform; HiLens, a developer-friendly AI application development platform; Atlas 200, an offline development kit which provides an online-equivalent offline experience; and HiQ, a quantum computing simulator and programming framework. With these complete platforms for end-to-end AI development, Huawei hopes to accelerate AI industry development with developers.
Developers face two major challenges in AI development. First, inefficient development, which makes labeling, training, and deployment extremely time-consuming. Second, the scarcity and high costs of basic AI resources, especially computing power. These two are interdependent as solving one is dependent on making significant improvements in the other.
The fact is that developers today require more computing power, yet AI computing power remains scarce and needs to be reused more. To help developers solve these two difficulties, Huawei Cloud has created four new products.
ModelArts for fast AI development
The ModelArts development platform is faster than any other AI development platform out there. It offers one-stop services for AI application development across the entire process, from data labeling and preparation, to model training, model tuning, and model deployment.
ModelArts is a full-pipeline platform that allows developers to get started faster, train faster, and deploy faster. With Huawei Cloud's faster AI platform, Huawei hopes to make AI more inclusive in collaboration with developers.
100x higher data labeling and preparation efficiency
In an investigation of the data labeling and preparation phase, we discovered that data labeling and preparation is highly time-consuming and labor intensive, accounting for about 50 percent of development time. ModelArts has a built-in AI data framework and uses an AI mechanism to manage data and iterative training to address the high volumes of labeled data. This is particularly useful for scenarios involving large volumes of data and can boost data labeling and preparation efficiency by 100-fold.
Time on model training halved
In the model training phase, ModelArts leverages various optimization techniques, in particular cascade hybrid parallel technology, to halve the time spent on model training for a given model, dataset, and set of hardware resources.
One-click model deployment to devices, edge, and cloud
In the model deployment phase, large-scale AI implementation and model deployment is highly complex. For example, smart transportation often requires the simultaneous, one-off deployment of updated models in line with various specifications of cameras from multiple vendors, which is highly time-consuming and labor-intensive. ModelArts enables one-click deployment of models to all edge and end devices. Online and batch reasoning for cloud deployment is also supported to meet the needs of various scenarios, including massive concurrency and distributed deployment.
Accelerating AI development and lowering entry barriers
In addition to improving labeling, training, and deployment, ModelArts also helps developers boost development efficiency. For AI to become prevalent across all industries, it must be an easy-to-use basic skill. ModelArts' automatic learning feature, including automatic design and automatic hyperparameter tuning, is designed to make it easy for all developers to use.
After training is completed with satisfactory results, it’s still often difficult to identify the raw data version. ModelArts provides end-to-end visualized management of the entire AI development lifecycle, from raw data and labeled data, to training operations, algorithms, models, and reasoning services. This platform automatically generates traceability maps for the management of millions of models, data sets, and services, without the user having to do anything. By selecting any model, users can trace the corresponding datasets, parameters, and deployment location of that model. Useful features such as resume training from checkpoint and compare training results are also possible, both of which are particularly popular with Huawei developers.
Helping developers build their own ecosystems
ModelArts provides a shared management platform that supports data, models, and APIs for developers.
Data and models can be shared internally across a business, helping enterprises improve AI development efficiency, build their own AI capabilities, and protect their AI information assets in a comprehensive way.
A model repository fosters open and healthy external ecosystems, helping AI developers monetize their knowledge and establish their own influence and ecosystems.
HiLens for developing vision AI
Vision is our primary sense, but the majority of developers lack skills in developing vision AI. This is why Huawei has launched HiLens, a developer-friendly vision development platform.
HiLens comprises an AI-capable camera and a cloud-based development platform. Skills can be formed on the platform through control code and models, with the AI models trained by ModelArts. HiLens is also compatible with other mainstream framework training models. Skills developed using HiLens can be directly deployed on any end device with an Ascend chip.
The AI-capable camera is a smart camera that supports reasoning. It’s designed for developers to develop vision applications for cloud-device synergy. HiLens comes with an embedded Ascend 310 chip, with a processing capacity of hundreds of frames a second and facial recognition within milliseconds. It also features built-in lightweight containers to minimize resource and network bandwidth use and enable fast downloads.
HiLens functions like a smart eye for AI applications, and is applicable to a wide range of application scenarios, including home robots, driver drowsiness detection, hazard detection, and logistics. Huawei hopes HiLens will inspire developers to create new innovative applications.
Atlas 200 for offline development with an online experience
Alongside the online full-pipeline AI development platform, Huawei also launched an offline development kit – Atlas 200 – to enable developers to conduct R&D offline. The Atlas 200 Developer Kit, while small, has powerful functions and an extensive range of interfaces. Embedded with Huawei's Ascend 310 chip, Atlas 200 supports reasoning and verification, as well as operator development.
Atlas 200 has a diverse built-in library for operators, supporting 80 percent of NN-class operators. It also features hybrid precision quantization tools based on low-bit, compression and sparsification, enabling optimal hardware platform selection for optimal energy efficiency ratio models. Developers can develop customized high-performance operators without having to understand complex compilation principles and hardware implementation.
Atlas 200 is compatible with all major frameworks. And original models can be quickly deployed on the development board for reasoning and verification without any additional development. Huawei has so far enabled a number of universities and enterprises to access Atlas 200.
HiQ for quantum computation simulators and programming frameworks
In preparation for the future, Huawei Cloud is exploring the world of quantum computing. Combining AI and quantum computing will bring huge computing power and new perspectives on algorithms. Quantum computing is a new type of qubit-based computing. It’s expected to bring exponential improvements and breakthroughs to many computing tasks. Current supercomputers would take about 150,000 years to factor a large 300-digit number. It would take a quantum computer a few seconds.
Although quantum computing has huge potential, the quantum chips required are still at the exploration stage, and many challenges lie ahead. Huawei Cloud's HiQ is a quantum computing simulator and programming framework. It not only features industry-leading full-amplitude and single-amplitude simulators, but also the industry's first cloud service for a high-performance quantum error correction circuit simulator that can simulate hundreds of thousands of qubit circuits. HiQ also incorporates the unique classical-quantum hybrid BlockUI.
Harnessing Huawei Cloud's strong computing power, HiQ can simulate circuits with 42 qubits for full amplitudes and 81 qubits for single amplitudes, with low-depth circuits able to simulate 169 qubits for single amplitudes. This makes it the industry's leading quantum circuit simulation cloud service. The HiQ cloud service platform will be open to the public to facilitate research and education in the quantum computing field.
With HiQ, Huawei hopes to empower academia and industry to jointly tackle challenges in quantum computing, advance the field, and train quantum computing engineers and quantum algorithm engineers.
With ModelArts, Atlas 200, HiLens, and HiQ, Huawei looks forward to working with developers to accelerate the large-scale adoption of AI across different industries, and make it affordable, effective, and reliable for all. Together with developers, Huawei is exploring the intelligent world of the future.