Connectionist Modeling of Part-Whole Analogy Learning

Analogical reasoning, along with inductive, deductive or abductive reasoning, belongs to the fundamental human mechanisms for the environment exploration, learning, or problem solving. Modeling this ability using computer simulations is important, as it might offer mechanistic explanation of these phenomena. In this work, we focus on the part–whole analogies in a separation task where the analogical objects between two scenes show a mutual resemblance. In simulations, using a simple recurrent network, we deal with the problem of geometrical analogies, inspired by the Analogator model (Blank, 1997). The original model had learning limitations hindering its full potential, combined with increased time and memory complexity. We propose model modifications for removing these limitations which leads to superior learning performance both in terms of speed and accuracy.