Transfer of Cognitive Skill: PRIMs
Is the whole of human cognitive ability an integrated system of knowledge, strategies and skills, or a collection of individual tasks and goals? Even though most people would gravitate towards the former point of view, the tradition of psychology, cognitive science and cognitive modeling adopts at least the research stance of the latter.
Formal discipline of the Mind or Identical Elements?
The discussion can be traced back to Thorndike (1902), who rejected the idea of a "formal discipline of the mind", and replaced it with the theory of identical elements. According to this theory, any ability we have is largely independent of other abilities, unless the two share identical elements of knowledge (Stimulus-Respons bonds in the time of Thorndike). Singley and Anderson (1985) introduced the modern version of this idea: transfer between individual skills is only possible if they share identical production rules.
Transfer, does it exist?
The current research tradition is to study individual skills and tasks with little regard for interactions between tasks. This is reinforced by many studies that demonstrate a lack of transfer, for example the well-known example in which subjects fail to solve a puzzle about a heart surgeon using a laser to remove a tumor after reading a story about a general who uses his army to concur the capital. Or the fact that even after taking a course in logic, students still fail to solve Wason's selection task.
From semantic to syntactic transfer
Most examples of failed transfer, however, play out on a semantic level, in which subjects fail to make the appropriate analogy, even if it is almost forced-fed into them. However, there are several experiments that do show transfer, but the transfer of knowledge seems to play out a more mechanical, syntactic level. In those experiments, subjects can perform or learn particular tasks faster because they have already learned a similar other task. Singley and Anderson's experiment of learning text editors is an example: it is easier to learn a new text editor if you have already mastered a different editor.
A new branch of more recent experiments have a similar structure, but focus on executive control. By training a particular control task, for example task switching, N-Back or working memory, subjects also improve on other executive control tasks, like the Stroop task.
PRIMs: A theory and model of syntactic transfer
To account for transfer, I have developed an extension to ACT-R that breaks down production rules into the smallest possible building blocks. Learning any skill leads to combinations of these building blocks, even though these combinations have no semantics of their own. They can be reused by other tasks and skills, and are therefore the basis for explaining transfer. PRIMs can explain transfer in Anderson and Singley's text editor study, and transfer in arithmetic. But PRIMs also offer explanations for the greater challenge: far transfer. In far transfer, the tasks involved have no surface characteristics in common, but nevertheless exhibit transfer. For example, training task switching leads to improvements on tests of working memory capacity and a decrease in Stroop interference. PRIMs can therefore help to evaluate the effectiveness of any kind of cognitive training that has the goal to improve general cognitive abilities.
Key references
Taatgen, N.A. (2013). The nature and transfer of cognitive skills. Psychological Review, 120(3), 439-471. (pdf)
Taatgen, N.A. (2013). Between architecture and model: Strategies for cognitive control. Biologically Inspired Cognitive Architectures. (link)
Taatgen, N.A. (2013). Diminishing return in transfer: A PRIM model of the Frensch (1991) arithmetic experiment. Proceedings of the 12th international conference on cognitive modeling (pp. 29-34). Ottawa: Cartelon University. (pdf)
Supplementary materials are on this page. This includes several models, documentation, and a version of ACT-R with which the models work.
How to run and learn PRIMs?
There are two options for running PRIMs. You can either load it as an extension to the ACT-R architecture (in Lisp), or you can download a standalone version that has been programmed in Swift. For the ACT-R extension you need Lisp, and for the standalone version you need a Mac. The new Mac implementation is strongly recommended if you start with PRIMs!
PRIMs implementation for MacOs
You can download a compiled version here. This version has the software, several example models, a tutorial document, and several powerpoint presentations. To make sure you have the latest version, go to https://github.com/ntaatgen/ACTransfer. In that case you will need to XCode to compile PRIMs.
The PRIMs tutorial is also on github. It contains a compiled version of PRIMs (for MacOS), a tutorial document, several papers, example models, exercises and Powerpoint presentations. The address is: https://github.com/ntaatgen/PRIMs-Tutorial
Last Lisp version of PRIMs
The new version of PRIMs has gradually deviated from the original Lisp version, so we do not recommend using it unless you have a clear goal. All the information below concerns the Lisp version, and should be considere outdated at this point.b>
The last versoion is version 0.55. Note that running this version will probably lead to faster (in ACT-R time) models after learning, because it allows parallel processing in places where they were forced to be serial in the previous version.
Running Actransfer
A zipfile with the Actransfer code and sample models.
A version of ACT-R (retrieved from the ACT-R website on June 28, 2012). Actransfer works with this version, but it may break future versions.
ICCM Paper
Taatgen, N.A. (2013). Diminishing return in transfer: a PRIM model of the Frensch (1991) arithmetic experiment. Submitted to ICCM 2013 (pdf). The model can be downloaded here. Note that to run the model the ACTransfer and ACT-R software needs to be downloaded and installed first using the links above. |
