Metacognition in the Age of AI: Challenges and Opportunities

Motivation

Recent progress in artificial intelligence and machine learning research has transformed the way we live, work and interact. Machines are mastering complex games and are learning increasingly challenging manipulation skills. Yet where are the robot agents that work for, with and alongside us? Recent successes heavily rely on the ability to learn at scale, often within the confines of a virtual environment, by trial and error over as many episodes as required. This presents significant challenges for embodied systems acting and interacting in the real world: an elaborate exploration of an agent’s state space is often unrealistic due to complexity and safety constraints; the critical inter-dependence of perception, planning and control coupled with limited hardware often leads to fragile performance and slow execution times; and cost of deployment severely limits the amount of training data obtainable. In contrast, we require our robots to robustly operate in real-time, to learn from a limited amount of data, take mission- and sometimes safety-critical decisions and increasingly even display a knack for creative problem solving. Achieving this goal will require artificial agents to be able to assess - or introspect - their own competencies and their understanding of the world.

Human psychology and cognitive science suggest that, while humans are faced with similar complexity, there are a number of mechanisms which allow us to successfully act and interact in the real world. Our ability to assess the quality of our own thinking - our capacity for metacognition - plays a central role. We posit that recent advances in machine learning have, for the first time, enabled the effective implementation and exploitation of similar processes in robotics. This workshop brings together experts from psychology and cognitive science with cutting-edge research in AI, robotics, representation learning and related disciplines with the ambitious aim of re-assessing how models of intelligence and metacognition can be leveraged in artificial agents given the potency of the toolset now available. With a particular focus on parallels in human metacognition, of particular interest will be cognitive models and neural mechanisms for metacognition applicable to AI systems, such as

1. computational models of metacognition for perception, planning and control;

2. architectures and implementations of metacognitive systems;

3. metacognition for robust, rapid or safe learning;

4. introspective or curious exploration, learning through interaction;

5. the applicability of dual process theory to artificial agents;

6. metacognition, world modelling and causal discovery;

7. metacognition for resilient action under model uncertainty and mis-specification;

8. datasets and metrics for evaluating the metacognitive capacity of artificial agents.