QToMNet: A Quantum-Enhanced Machine Theory of Mind

Theory of Mind (ToM) broadly refers to a human’s cognitive capacity to attribute mental states, such as beliefs, to oneself and others. The emerging stream of machine ToM seeks to equip artificial agents with this ability using machine learning, including neural networks. In this paper, we build upon evidence on the quantum dynamics of human decision making and hypothesize that quantum embeddings are a promising direction for enhancing machine ToM neural networks, such as ToMNet and its extensions. Such meta-learning models learn a small number of behavioral observations to make rich predictions about other agents’ characteristics and mental states. Quantum embeddings map outputs of ToMNet’s character and mental networks into parameterized superimposed and entangled quantum states, and provide a mathematical structure that allows the network to better approximate the uncertainty inherent to partially observable mental states. We outline a quantum-enhanced version of ToMNet, called QToMNet, which uses quantum embeddings to encode past episodes and current trajectories into quantum circuits, with a classical prediction head producing standard ToMNet outputs. Further, we present an experimental plan to evaluate QToMNet’s predictive power using grid-world settings with random and deep reinforcement learning (RL) agents. With this position paper, we aim to catalyze rigorous experiments that examine whether quantum embeddings can offer practical, drop-in gains in performance of ToM algorithms, and hope to nurture a new stream of literature on quantum-enhanced ToM networks.