People usually sense when they step onto unfamiliar ground and their knowledge runs out. Artificial intelligence has no such instinct: it learns from selected data and can be helpless when the world looks different from training. The European E-pi project taught algorithms to recognise their own ignorance and checked it where a mistake can cost the most, in autonomous cars. At the end you can find out how you handle uncertainty yourself.
Artificial intelligence now reaches almost everywhere. It drives the economy and is changing our homes, schools and workplaces. Yet it has a weak spot that is mentioned less often than its successes. A model learns from a specific set of data and then has to work in a world that does not always resemble what it saw during training. That is exactly when the trouble begins.
The EU-funded E-pi project, coordinated by Oxford Brookes University in the United Kingdom, took on this problem. Over five years its team worked on so-called epistemic artificial intelligence, meaning AI that copes better with uncertainty and can make predictions even in situations that differ from its training data.
The word epistemic comes from epistemology, the branch of philosophy that studies knowledge. From this field comes an ability that is completely natural for a human: noticing that what we know is not enough in a given situation. It is one of the basic elements of human intelligence.
A machine cannot do this on its own. For AI, recognising its own ignorance has to be built on theoretical and mathematical frameworks. In other words, the feeling of not being sure any more has to be written, quite literally, into the algorithm’s design.
Project coordinator Fabio Cuzzolin points out that classic machine learning algorithms focus on explaining a relatively small training set rather than all the data that may appear in reality. The consequence is easy to predict. “Traditional AI models struggle to make robust predictions when the probability distributions generating the data differ from those at training time,” the researcher explains.
Put simply, a model that knows only a slice of the world well can get lost when it meets something outside that slice. The E-pi consortium concluded that second-order uncertainty theory is the most promising foundation for a solution. The project’s presentations were received with enthusiasm at professional conferences, and Cuzzolin was invited to take part in several related Horizon Europe proposals. The team’s work is meant to help build an international ecosystem of epistemic AI.
Where is the best place to test an AI that is supposed to cope with the unpredictable? The researchers chose autonomous vehicles. On the road a situation can suddenly break out of any pattern, whether through the behaviour of pedestrians, extreme weather or an unusual road layout. The mistakes of current systems can be costly and, in the worst case, may result in fatal injuries.
The team used the ROAD dataset, devoted to the recognition of road events by autonomous vehicles. It served both to train the model and to test new epistemic object detectors. ROAD keeps being expanded and contains videos from Europe, North America, Dubai and Abu Dhabi, which has made it a truly global resource.
Safer driverless cars are just one goal. Other fields can benefit from the same logic too:
Nuclear fusion is an especially demanding example. “We are working to develop uncertainty-aware neural operator models to predict how the fusion plasma will flow within a tokamak and which areas of the reactors will be more affected by it,” says Cuzzolin. Such models are meant to forecast the flow of plasma inside a tokamak, that is, a reactor, and to indicate which of its areas will be most exposed to it. Here too the point is the same: a model should not only give a result but also honestly account for how sure it is of it.
Another direction concerns generative AI. “Current state-of-the-art large language models (LLMs) are known for generating false statements,” Cuzzolin admits. In his view, the way out may be models built on a new paradigm: “Epistemic LLMs can significantly mitigate the overconfidence of those models.”
The plans also include solutions related to theory of mind and systems that evolve through continual learning. The common denominator stays the same: artificial intelligence should not only answer but also know when its answer rests on shaky foundations.
The E-pi project, whose total cost was about €3 million, contributed to the development of safer autonomous vehicles and robots, and of generative artificial intelligence that does not hallucinate. It also built a dynamic network of machine learning experts. Its authors argue that epistemic AI has the potential to change the world and our everyday lives. And it all begins with a very human skill: admitting that you do not know something.
5 questions · one minute
English
Čeština
Deutsch
Español
Français
Magyar
Italiano
Polski
Português
Русский
Slovenčina
Türkçe
Українська
العربية
中文