Problem/ Challenge:

More and more AI technologies are being transferred from research to practical application – whether in the private or in the public sector. One of the biggest challenges in this transition to the real utilisation of artificial intelligence is the underlying data. Only if suitable data is utilised for the development and evaluation of AI models will these models be reliable and efficient. If this is not the case, the growing number of AI initiatives by companies and public organisations will fail due to unexpected expenses, missed deadlines or regulatory risks.

The data-related challenges in the real-world use of AI are manifold: they range from poor data planning and inadequate data descriptions to the multiple development of data-related solutions for the same problem. Another difficulty lies in dealing with synthetic (artificially produced) data. Inadequate synthetic data sets, for example, harbour the problem that they can show distorted results or be inaccurate. Tracking down the causes of this is highly complex and requires particular attention.

Furthermore, the development of AI models for a laboratory environment differs drastically from the development of models that have to work in the real world – for example in industrial production. In addition to fine-tuning algorithms, it is important to ensure that the data sets that are used are actually suitable for the task at hand. At the same time, the traceability of AI continues to pose high demands: Customised explanations are needed through approaches that enable as many users as possible to understand and trust AI predictions.

Approach/ Solution:

The EU project AI-Dapt pursues a data-centred AI approach with the objective of strengthening the trustworthiness of AI solutions for their use in real-world environments. Under the leadership of the “Athena Resea rch Centre”, 18 project partners are researching on hybrid AI solutions that combine data-driven AI with scientific and theoretical expertise.

To this end, the project partners develop data-centred solutions for the entire life cycle of the deployment of artificial intelligence in productive environments. These enable an end-to-end automation across the entire cycle – from design, execution and evaluation to the optimisation of intelligent data AI pipelines that continuously learn and adapt to their context. The pipelines include steps such as data acquisition, data cleansing and data analysis as well as the development and training of AI models and their deployment and updating.

AI-DAPT solutions also enable continuous, dynamic improvements to AI during operations: for example, the training of AI models can be adapted without interruption or the time from problem detection to solution can be shortened. The overarching objective is to ensure that artificial intelligence works reliably in different production environments.

The AI-DAPT services:

Specifically, AI-DAPT develops and links a series of scalable and interoperable services that are applied within a specially developed AI-DAPT platform – and can also be integrated into other data analysis platforms. These services are organised along the following five axes:

• Data design for AI, e.g. data analysis, mining, harvesting and evaluation
• Data maintenance for AI, e.g. data cleansing, enrichment and selection
• Data generation for AI, e.g. evaluation of data usability and generation of synthetic data
• Provision of models for AI, e.g. model training and evaluation
• Data model optimisation for AI, e.g. data and model monitoring

AI-DAPT will combine automation across all five axes with targeted “human-in-the-loop” approaches. These ensure that humans play an active role in the AI system – for example, by checking data or taking decisions suggested by the AI. This involves the deployment of XAI (explainable AI) technologies: they make the AI's decisions understandable for users and ensure the ethical use of the AI in operation.

The AI-DAPT solutions will be tested and continuously reviewed under real-life conditions by demonstrators in four different sectors: healthcare, robotics, energy and manufacturing. Together with project partners, Fraunhofer FOKUS supports and develops services for the design and maintenance of the data, assists in the architectural development of the pipeline and develops standards for communication between the services.