Although artificial intelligence (AI) has improved remarkably over the last years, its inability to deal with fundamental uncertainty severely limits its application. This proposal re-imagines AI with a proper treatment of the uncertainty stemming from our forcibly partial knowledge of the world.

As currently practiced, AI cannot confidently make predictions robust enough to stand the test of data generated by processes different (even by tiny details, as shown by ‘adversarial’ results able to fool deep neural networks) from those studied at training time. While recognising this issue under different names (e.g. ‘overfitting’), traditional machine learning seems unable to address it in nonincremental ways. As a result, AI systems suffer from brittle behaviour, and find difficult to operate in new situations, e.g. adapting to driving in heavy rain or to other road users’ different styles of driving, e.g. deriving from cultural traits.

Epistemic AI’s overall objective is to create a new paradigm for a next-generation artificial intelligence providing worst-case guarantees on its predictions thanks to a proper modelling of real-world uncertainties.

Epistemic AI (E-pi) is a research project funded by the European Union under its Horizon 2020 initiative, in particular the Future Emerging Technologies (FET) scheme which is designed to foster blue-sky thinking and bring about paradigm changes.

This is a 4-year (48 months) project which started on 1 March 2021, with a budget of around 3 million euros and three international partners (all universities): Oxford Brookes University (United Kingdom), KU Leuven (Belgium) and Delft University of Technology (TU Delft, Netherlands). 

Epistemic AI breaks entirely with the current state of artificial intelligence and with the most exciting ongoing efforts, such as continual learning (making the learning process a life-long endeavour), multi-task learning (aiming to distil knowledge from multiple tasks to solve a different problem) or meta-learning (learning to learn). As these are all still firmly rooted in AI’s conventional principles, they fail to recognise the foundational issue that the discipline has with the representation of uncertain knowledge.

Our proposal goes beyond ‘human-centric’ AI, the push to make artificial constructs more trustable by human beings and more capable of understanding humans, since it strives to model the uncertainty stemming not just from human behaviour, but from all sources of uncertainty present in complex environments.

Epistemic AI’s overall objective is to create a new paradigm for a next-generation artificial intelligence providing worst-case guarantees on its predictions, thanks to a proper modelling of real-world uncertainties. Firstly, a new mathematical framework for optimisation under epistemic uncertainty will be formulated, superseding existing probabilistic approaches. The new optimisation framework will lay the premises for the creation of new ‘epistemic’ learning paradigms. In Epistemic AI we will focus, in particular, on some of the most important areas of machine learning: unsupervised learning, supervised learning and reinforcement learning.

Last but not least, the goal of the project is to foster an ecosystem of academic, research, industry and societal partners throughout Europe able to drive and sustain the EU’s leadership ambition in the search for a next-generation AI.

While traditional ML learns from the (limited) available evidence a model able to describe it, with limited power of generalisation (see the figure below, left), epistemic AI (right) starts by assuming that the task at hand is (almost) completely unknown, because of the sheer imbalance between what we know and what we do not know. Anything is actually possible. Our ignorance is only tempered in the light of the (limited) available evidence to avoid ‘catastrophically forgetting’, to use a trendy expression, how much we ignore about the problem or, in fact, that we even ignore how much we ignore. Mathematically, Epistemic AI’s principle translates into seeking to learn sets of hypotheses compatible with the (scarce) data available, rather than individual models. A set of models can provide, given new data, a robust set of predictions among which the most cautious one can be adopted, thus avoiding catastrophic results.