Causal AI: The future of machine learning

Sunlight streams through a large window into Francesco Locatello's new office. From the campus panorama, your gaze immediately falls on the pictures in the room. The photographs are barely recognisable. Only upon closer inspection do they become clearer. A small green blob and a star-shaped pink scribble. Slowly, it dawns on you. ‘The green picture is a cabbage and the other shows starfish,’ Locatello explains. ‘Both look realistic, but somehow they don't.’ The colourful prints are AI art by Tom White, whose series pays homage to how machines perceive our world. This is precisely where Locatello's research comes in. ‘Among other things, we are trying to help machines perceive the world as we do,’ says the computer scientist.

Correlation is not causation

Locatello's research focuses on machine learning and artificial intelligence (AI), two topics that have seen significant advances, particularly in the last five years. We are now surrounded by a range of different AI models that make our lives easier: they recognise images, reproduce natural languages, process knowledge, or are built into self-driving cars that drive up and down the hills of San Francisco. However, current machine learning mostly only scratches the surface of reality. It only works perfectly under consistent conditions, and changes in the real world, changing settings, or the development of circumstances over time are largely ignored.

In order to push the boundaries of machine learning, Locatello's group is focusing on learning causal representations and models from large amounts of data. ‘We are particularly fascinated by AI systems that are able to “understand” causal relationships between events – in other words, why a certain action leads to a certain result,’ says Locatello. To this end, the researchers are developing theories and scalable algorithms that will enable AI agents to grasp cause-and-effect relationships, i.e. when one thing leads to another (think of falling dominoes or a Newton's cradle). 

‘As AI technologies become increasingly integrated into our daily lives, it is important to understand what would happen if we actively intervened in these systems,’ explains the computer scientist. ‘In the long term, we want to design AI systems in such a way that they can help us generate new knowledge about the world. And in a way, that's not so far-fetched.’

Locatello is convinced that the interdisciplinary nature of ISTA will be helpful in this regard. ‘The main reason I chose ISTA was the broad spectrum of researchers from different disciplines. It would be fantastic if we could train AI models that ultimately discover something about astronomy, the universe or the climate!’

A scientific career from Venice to Vienna

Locatello grew up in a small town near Venice and has always been interested in computers. However, his love for AI technologies, research and machines only developed when he was a doctoral student at ETH Zurich and the Max Planck Institute for Intelligent Systems from 2016 to 2020. During this period, he worked with Google as a research consultant and as an intern in the software department. After graduating, he became a senior applied scientist and led his own research team at Amazon Web Services, focusing on causal representation learning.

He comes to ISTA with several awards under his belt – the Best Paper Award at the International Conference on Machine Learning (ICML) 2019 and the Hector Foundation Prize for outstanding scientific achievements in the field of machine learning.

To bring a touch of Italy to the office all year round, it is adorned with a small olive tree and a lemon tree. The lemon tree even bears bright yellow fruit – presumably an omen of a successful future on campus.