A spiking neural model applied to the study of human performance and cognitive decline on Raven's Advanced Progressive Matrices

Abstract

We present a spiking neural model capable of solving a popular test of intelligence, Raven's Advanced Progressive Matrices (RPM). The central features of this model are its ability to dynamically generate the rules needed to solve the RPM and its biologically detailed implementation in spiking neurons. We describe the rule generation processes, and demonstrate the model's ability to use the resulting rules to solve the RPM with similar performance and error patterns to human subjects. Investigating the rules in more detail, we show that they successfully capture abstract patterns in the data, enabling them to generalize to novel matrices. We also show that the same model can be used to solve a separate reasoning task, and demonstrates the expected positive correlation in performance across tasks. Finally, we demonstrate the advantages of the biologically detailed implementation by using the model to connect behavioral and neurophysiological data. Specifically, we investigate two neurophysiological explanations of cognitive decline in aging: neuron loss and representational “dedifferentiation”. We show that manipulations to the model that reflect these neurophysiological hypotheses result in performance changes that match observed human behavioral data.