Modelling Memory for WHERE
We had participants collect audio, accelerometry, GPS and emotion data continually over a two week period. A week later, we presented four alternative forced choice trials in which they had to identify where they had been at given times. To model the data, we created a memory network for each individual which contained nodes representing GPS, accelerometry and audio clusters as well as time (hour, day, week) and discrete emotions. Links were added between nodes that co-occurred most frequently with a constraint on the maximum degree. Each trial was modelled by adding the outer product of the time cue vector with itself to the memory network (simulating short term plasticity) and then cycling activations to find the primary eigenvector. The dynamic eigen network model with max degree learning rule captured 51.3% of participant choices (well above chance) exemplifying generalization at retrieval (c.f. Hintzmann, 1988) through the dynamic reconfiguration of the eigenstructure of the memory network.
Cite this as:
Dennis, S., Yim, H., Shabahang, K., Laliberte, E., & McKenzie, A. (2021, February).