Safer systems through probabilistic machine learning

Artificial Intelligence (AI) and data-driven decisions based on Machine Learning (ML) are making an impact on an increasing number of industries. As these autonomous and self-learning systems become more and more responsible for making decisions that may ultimately affect the safety of personnel, assets, or the environment, the need to ensure the safe use of AI in systems will be crucial to safety management and operations.

Machine learning for high-risk and safety-critical applications is challenging, as there is a reduced tolerance for erroneous predictions due to potentially catastrophic consequences. The models which fit the data well are also often opaque, making them less falsifiable and difficult to trust. Moreover, there is usually a dearth of relevant data, and a proper treatment of uncertainty is essential, as DNV is not only concerned with what is likely to happen, but also with less likely events that may happen. However, there are some positives – namely, that there is often additional causal and physics-based knowledge available.

Probabilistic machine learning models

The project will develop practical algorithms combining information about causal relationships from phenomenological knowledge (what depends on what), with information on correlation (how much) from data, together with the theoretical results needed for safe implementation of AI/ML systems in a high-risk environment.

It will help achieve this by combining causality constraints derived from phenomenological knowledge with probabilistic machine learning. For example, using criteria such as “stronger material implies higher capacity” to obtain a machine learning model that adheres to physical laws. It will also make use of synthetic (simulation) data to close the gaps in existing datasets, while building in causal dependencies by design, using the classical expert systems approach to AI in combination with the modern ML based approach.

This will hinge on developments in active learning, through ML systems which self-improve by asking the right questions or perform optimal experiments. This is essential when learning from a limited set of observations, and for use in applications that involve optimal decision making under uncertainty.

The benefits

Most applications of AI/ML today are related to low consequence scenarios, but the project aims to develop a framework suitable for safe implementation in a high-risk environment. DNV focuses, in particular, on how phenomenological knowledge can be incorporated in probabilistic machine learning models, together with the use of active learning for optimal decision making under uncertainty.

DNV believes that AI/Machine Learning will play an important role in safety-critical systems in the future, and the capability of handling this in a safe manner provides significant market potential.