The crucial role of Theory of Mind (ToM) capabilities in developing a next generation, human-centric artificial intelligence

Emerging applications of artificial intelligence (AI) are raising awareness of the limitations of established machine learning (ML) approaches in situations in which humans are involved. Smart cars, for instance, need to make reliable predictions about human behaviour in real time, e.g. in order to pre-emptively adjust speed and course to cope with a group of children’s possible decision to suddenly cross the road in front of them. To date, the automated recognition of present behaviour builds on the recent successes of deep learning (a technique based on artificial neural networks with an increased number of layers) to efficiently identify motion patterns in the available streaming videos. Motion patterns, however, can be deceiving, as humans can suddenly change their mind based on their own internal mental dynamics and things they spot in the scene (e.g., children seeing an ice cream van on the other side of the road). Intriguingly, humans are capable of making reliable predictions of future behaviour even when no motion is present, just by quickly assessing the ‘type’ of person they are looking at (e.g. an elderly person standing in a hallway is likely to take the elevator rather than the stairs). Theory of Mind (ToM) capabilities, i.e., the ability to ‘read’ other agent’s mental states, are crucial to develop a next generation, human-centric artificial intelligence. In a mutually beneficial process, computational models developed within AI may provide new insight on the way these mechanisms work in the human brain. Within cognitive neuroscience, the theory-theory paradigm argues in favour of the existence of a set of rules humans possess regarding human mind functioning. The simulation-theory view defends, instead, a simulation process consisting of taking someone else’s perspective to understand their reasoning. The simulation standpoint, we argue, has the potential to bring together machine learning and neuroscience in a radically new approach to the problem. A fruitful cross-fertilisation of neuroscience and machine learning can enable significant advances in both fields, by allowing both the formulation of computational theory of mind models in humans leveraging current frontier efforts in AI, and the development of machine theory of mind models informed by the most recent neuroscientific evidence, capable of going beyond simple pattern recognition for prediction in complex, human-centred scenarios. ‘Belief-desire-intention’ models are so far the dominant approach in computational ToM. However, as they lack the ability to learn from past experiences, such methods have been subject to much criticism. In opposition, composable deep networks have the potential to provide solid foundations for a simulation-theory approach. The reasoning of various classes of complex agents can be flexibly simulated by dynamically rearranging the topology of the connections among a base set of neural modules. The way the simulation adapts to agent class and scene is learned from experience via reinforcement learning. Successful machine Theory of Mind models would lay the foundations for the creation, among others, of autonomous vehicles able to negotiate complex road situations involving humans. Next-generation robotic assistant surgeons can be envisaged with the ability to understand what the main surgeon is doing and foresee their future intentions. Empathic robotic healthcare tools would become possible, as well as a new generation of ‘bots’ (e.g. for customer service or financial advice) able to interact more effectively and empathetically with humans. Research in this area would also impact on the current debate on moral AI, helping machines make ethical, human-like decisions in critical situations.