PRIMs Cognitive Architecture

Transfer Learning

People typically do not learn new things in isolation, and therefore any new task we learn will be affected by our prior learning history. Although this may sounds as something completely obvious, cognitive architectures typically do study tasks and task knowledge in isolation.

If we want to study how prior experience with one task affects performance and learning on another task, we talk about Transfer Learning. PRIMs has several example models of transfer. The most "classic" one is a model of a text editor experiment performed by Singley and Anderson in which subjects were trained on one text editor, and then had to switch to another.

Models made by Singley and Anderson were not completely capable of explaining all of the transfer effects in the data, in particular that prior training on a line editor transfers for 63% to the Emacs screen-based editor. The PRIMs model broke down the knowledge for the editors in smaller basic operators than the previous approaches, and was able to show that strings of "PRIMs" were identical in many places in the model (see graph: each node is a PRIM). Overlap between the two line editors in the experiment was largest, but there was also substantial overlap between the Emacs editor and the two line editors.

See for more details (Taatgen, 2013a).