Combining Mental Models with Neural Networks

Human reasoning under uncertainty is conjectured to use Mental Models as a representation format. Each Mental Model characterizes a possible state of the world based on and constrained by the available information. Conclusion about the world must hold in each of these Mental Models. An important task in human reasoning is the construction these Mental Models using the available information. This paper investigates whether it is possible to design a neural network architecture that enables the construction of Mental Model, similarly to the conjectured way that humans reason. The paper investigates different architectures in an incremental way. The final architecture not only produces the correct mental models but also learns correct mental models for intermediate representation without being explicitly trained to do so. This contributes to the explainability of the approach.