Thought experiments combined with physics simulations and machine learning in a human-AI learning context

Using thought experiments as a context for creating and exploring a novel artificial intelligence (AI) system that captures dynamics of human cognition in the form of mental models. A representation and prediction system is trained to emulate the basic causal chain-of events occurring when humans engage in thought experiments. Basic operations are cast into simple dynamical rules and objects, defined in a similar way as in cellular automata and inspired by production rules in symbolic cognitive architectures, using an autoencoder-like architecture for enhanced abstraction. Potential outcomes of thought experiments can thus be “played out”, generated by the physical dynamics, and iteratively refined. These “AI augmented” thought experiments provide a direct means for self-reflection and critique of one’s mental models (i.e., unconscious biases and presumptions), and stimulate a creative process. At the same time, this enables building an AI model consistent with our own understanding. Cross-disciplinary implications are discussed upon evaluating the system from the perspectives of cognitive science, physics of non-equilibrium, complex systems, as well as for AI performative aspects like knowledge transfer.