No description, website, or topics provided.
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
main code
README.md
Simulation1.m
Simulation1_plots.m
Simulation2.m
Simulation2_plots.m

README.md

CogSci-2017

Source code for analyses reported in

Musslick, S., Saxe, M. A., Özcimder K., Dey, B., Henselman, G., Cohen, J. D. (2017). Multitasking Capability Versus Learning Efficiency in Neural Network Architectures. Proceedings of the 39th Annual Meeting of the Cognitive Science Society.

Simulation 1

The simulation involves training a neural network on a set of tasks in an environments with different task similarities. Each task requires to map a particular set of 5 stimulus features to a particular set of 5 response units. Task similarity between a pair of two tasks is operationalized in terms of the number of shared relevant stimulus features in the environment (feature overlap). At each training iteration the learning performance, as well as multitasking performance are assessed.

To generate a replication of the stimulation index by 'rep', run Simulation1(rep)

E.g. Simulation1(1)

Results averaged across multiple replications can be plotted using the code in Simulation1_plots.m

Simulation 2

The simulation involves training a neural networks on a set of tasks with different initial weight conditions. Each initial weight condition determines the initial similarity between tasks that rely on the same input dimension. All tasks are trained in an environment with a fixed amount of feature overlap. At each training iteration the learning performance, as well as multitasking performance are assessed.

To generate a replication of the stimulation index by 'rep' in an environment with a particular feature overlap FO (can range from 0 to 1 in steps of 0.2), run Simulation2(FO, rep)

E.g. Simulation2(1.0, 1)

Results averaged across multiple replications can be plotted using the code in Simulation2_plots.m