This is our time: Working on reliable AI

Friso Heslinga conducts research on the use of AI to make the best possible decisions. ‘I develop AI models that can automatically assess photo and video images. Cameras, satellites, self-driving cars, hospitals, businesses – we are capturing more and more images, sometimes 24/7. Interpreting all those photos and videos is almost impossible without AI.’

Learning from examples

The more information AI can interpret, the better and faster people can make decisions. ‘An AI model learns from examples, which we call a data-driven approach. This way, the model learns to detect anomalies on its own.’

From diagnosing diseases to a car recognising its surroundings, image analysis with AI can add value in various ways. During his PhD, Friso focused on medical image analysis, including supporting corneal transplants. Additionally, Friso developed AI for the analysis of retinal images, a subject where TNO has also made significant strides.

When data is lacking

How do you train an AI model to recognise something when there are hardly any relevant images and videos available? To solve this, Friso is exploring several possibilities at TNO, one of which is supplementing real data with synthetic data. ‘Synthetic data is data we create ourselves using 3D simulation software. Recently, we have also been creating this data with generative AI. For example we make existing synthetic datasets even more realistic or add more variation so that the AI model learns to recognise small differences.’

Trusting AI

People need to be able to trust the AI models that support them. This is a major challenge, Friso explains. ‘People are very good at generalising. Once they have seen a few examples, they can apply their knowledge to new situations, even if they look quite different. Generalising is a big challenge for AI models, especially if they are trained with synthetic data. We still need to improve our simulations and AI models. The rise of generative AI offers enormous opportunities, for example, by making already simulated images extra realistic. Or even better, by directly converting a textual description into a realistic image.’

Accelerating change

AI models were partly inspired by the neural networks of the human brain. Many steps have since been taken in terms of model architectures and training methods. Friso: ‘The technology and speed of developments continue to fascinate me. AI can combine more and more different tasks. This is already changing how we work, and these changes will soon accelerate even further. The impact of AI will be very significant, in all aspects of our lives ultimately. Being able to contribute to this is what make my work really cool.’