The Future of AI

The current success of Artificial Intelligence can be attributed to machine learning methods that can generate predictions on the basis of huge volumes of data. The Bernoulli Institute is focused on the next generation of AI.

Persistent human-inspired artificial intelligence. Natural intelligent systems are not trained and then tested, but exist in a changing world that requires continuous adaptation and learning. To build AI systems with these capabilities, a number of hard problems have to be solved. In particular, AI systems should be able to build new knowledge on top of the knowledge they already have, and use existing skills to solve new problems. Solving this problem would allow us to build AI systems that no longer require the vast amounts of data that are now needed, and make them more robust against unexpected and novel contexts.

Human-centered AI. Developments in AI should benefit society, and not pose a threat. To be able to serve as a useful assistant, an AI system should be reliable and should be able to explain itself. In addition, the interaction itself should be a topic of research.

Hardware for AI. Current computer hardware is far from optimal to simulate neural networks. New neuromorphic hardware can help build AI-based computers of the future. This new hardware requires rethinking of the entire computer architecture “stack”, including analog computing, learning and transfer at different levels of abstraction, and the use of natural language.