Description

Cognitive maps represent the spatial, temporal, and conceptual relations underlying events occuring in an environment. Recently, it has been proposed that the computational structure of a cognitive map might correspond to a clone-structured cognitive graph (CSCG; George et al., 2021). One important property of cognitive maps is that they contain representations for flexible behavior, so that the system can efficiently learn and deal with ambiguous situations e.g. when similar observations occur in different contexts. Learning in CSCGs can be implemented via Expectation Maximization in Hidden Markov Models using a modified Baum-Welch algorithm. In this project you will study whether CSCGs are suitable to describe learning situations that humans face, where often only a limited number of training trials are available. To do so you will implement and study an adaptive, online version of the CSCG learning algorithm in sequence learning tasks, a simplified account for episodic memory in humans. Furthermore, you will study how key events in sequences with overlapping observations trigger the recruitment of novel nodes (“cloning”) and how this relates to prediction error based forms of structural learning (Gershman et al., 2017). This is important in the context of how artificial systems can be equipped with cognitive components that allow them to perform complex tasks.

References / Literature

Gershman, S. J., Monfils, M. H., Norman, K. A., & Niv, Y. (2017). The computational nature of memory modification. Elife, 6, e23763.

George, D., Rikhye, R. V., Gothoskar, N., Guntupalli, J. S., Dedieu, A., & Lázaro-Gredilla, M. (2021). Clone-structured graph representations enable flexible learning and vicarious evaluation of cognitive maps. Nature communications, 12(1), 2392.