In this article you will read about,

• how you can use ModelArts to develop AI applications – with or without programming skills
• which frameworks and tools ModelArts provides for professional developers
• how the end-to-end development platform can save time, effort, and costs

Artificial intelligence (AI) is increasingly becoming a competitive factor: 79 percent of the German companies that took part in a worldwide Deloitte study already consider artificial intelligence to be very important or a critical factor for success. 75 percent of the companies in Germany surveyed by the Fraunhofer IAO are currently working on questions relating to artificial intelligence, 16 percent are already using it – and say that they greatly benefit from their respective AI applications.

The ModelArts development environment is now available in the Open Telekom Cloud to develop AI applications. It allows not only experienced developers, but also users without programming skills to train and deploy AI models – thanks to an easy-to-use user interface and an auto learning function. Naturally, in addition to auto learning, ModelArts also supports machine learning and deep learning.

No AI algorithm can be trained without data – that's why ModelArts also supports data preparation and data annotation, which take up about 80 percent of an AI's training time. The developers use object storage to bring their data to the platform, where they can label it and store it efficiently using version management.

For those who have no experience in training algorithms and don’t know any programming language, the next step is to access pre-trained models that can be applied to their own data. Object and image recognition as well as predictive analytics are available, which are used, for example, to calculate maintenance intervals. For object and image recognition, images of the respective object are first imported into the program, whereupon the AI algorithm begins to train the recognition of objects and their exact position in the image (object detection). Once the training has been completed on the basis of the available data, the finished model can be made available at the click of a mouse, so that the skills learned by the AI can be tested immediately in practice on unknown images. Alternatively, it is possible to continue the training with additional data. A trained AI application is not only able to recognize objects, but also to assign them to a certain category and, for example, to distinguish between defective and intact screws (image classification).

For users with some experience in coding, ModelArts provides pre-configured algorithms for the training phase. In addition, experienced developers can use pre-installed programming frameworks such as TensorFlow – other popular open source software such as Pytorch, MXNet, Caffe, SparkML, Scikit, XGBoost and the MoXing Software Development Kit (SDK) are also available for experts. Those who want to work with their own frameworks and algorithms can easily integrate them into the ModelArts development environment via containers.

Last but not least, the open source web application Jupyter Notebook can be used as a toolbox that is started as a virtual machine in the Open Telekom Cloud and provides access to the required computing and storage capacity. With the help of Jupyter Notebook, developers can code and simultaneously test, visualize their work, and present it in reports. In this way, ModelArts offers the functionalities and tools that are suited to each developer’s skillset and programming experience.

After the training is completed, the finished AI model can be deployed as an online service in the cloud, and can be accessed by other programs via an application programming interface (API). It is also possible to use the AI service as a batch job and, for example, have it run and compute overnight. ModelArts will soon also support edge devices so that industrial robots, for example, can use a fully trained AI application.

The most suitable processors are used depending on the application: CPUs (central processing units), GPUs (graphics processing units) or NPUs (neural processing units). Nvidia V100 GPUs are considered the best choice for image recognition. If, for example, a finished model for image recognition is provided as an online service via API, a server equipped with GPU chips is set up in the background, where the AI service can be accessed.

ith the ModelArts end-to-end platform from the Open Telekom Cloud, practically any user can train algorithms and start the finished AI models with a simple click of the mouse – regardless of prior knowledge or programming experience. ModelArts guides AI novices safely through the entire development phase with a complete guide and process description. More experienced developers can use the toolbox, which is equipped with numerous frameworks, to develop any kind of AI application case. All users benefit from the savings in time and effort that result from using existing components. Thus, the path to the finished AI model is considerably shortened – and all the project steps are documented by a transparent end-to-end lifecycle management

In addition, the Open Telekom Cloud offers all ModelArts developers virtually unlimited computing resources – a clear cost advantage over an in-house data center that has to be constantly upgraded with new hardware such as GPUs. In the Open Telekom Cloud, developers always use the latest generation of processors at demand-based rates. NPU chips will also be available from early 2021.

Use all services from one provider

Telekom supports AI developers with services from one source, from data preparation and training to the provision of the application. Processing in German data centers ensures the highest level of data protection and data security. Interested parties can access ModelArts via the Open Telekom Cloud website, which also offers additional AI services.