Neural network model of semantic processing in the remote associates test

The ability to generate novel, unique and useful ideas is an important trait of intelligent behaviour. It is also a virtue of a creative individual in many scientific and artistic domains. In this study we are concerned with the Remote Associates Test (RAT), a task widely used in psychology and neuroscience to study insight and creative problem solving. The RAT is used to assess the ability of an individual to generate novel relationships among familiar words. The test consists of word triplets (e.g. cream, water, skate) and the task is to find a unique word associated with all three words. Here, we aim to identify a basic set of computational mechanisms underlying cognitive processes in the RAT solving. To this end, we propose a multi-layer neural network based on biologically and cognitive realistic mechanisms. The search for a solution in a RAT problem is realised by spreading of activity among word associations in a semantic layer, and the selection of a response by a winner-take-all layer. The model yields human-like performance and distinguishes between easy and difficult RAT problems. The modelling findings are consistent with the existing theories in creativity research, confirming that less stereotypical word associations are important for the good performance on the RAT.