Learning proper symbol sequencing as an instance of structural abstraction

19/03/2021, 16:00 (GMT+1, Warsaw), Zoom.

This Friday, 19.03, 16:00, at the HILL seminar there will be a presentation given by Maksymilian SmelczyƄski.

Learning proper symbol sequencing as an instance of structural abstraction

Abstract Recently there has been a renewed interest in the hypothesis that cognitive structures supporting abstract thinking and spatial cognition may share common mechanisms (Constantinescu et al., 2016; Bellmund et al., 2018). Whittington et al. (2020) proposed a computational model of the hippocampal formation, unifying its involvement in spatial and relational memory, by re-casting both functions as examples of structural abstraction. In this work, we apply this idea to the problem of learning simple grammatical structures, inspired by the experiments conducted by Savage-Rumbaugh and Rumbaugh (1978). They aimed at teaching chimpanzees to properly sequence symbols in order to acquire rewards. By training a neural network with the abovementioned architecture on tasks that faithfully recreate these experiments, we attempt to show how spatial and abstract relations may be learned in a single system and aid the generalization of grammatical structures to previously unseen symbols.

Recommended reading: https://science.sciencemag.org/content/362/6415/eaat6766 (Bellmund et al. 2018)