Flexible task abstractions emerge in linear networks with fast and bounded units

by Kai J. Sandbrink*, Jan P. Bauer*, Alexandra M. Proca*, Andrew M. Saxe, Christopher Summerfield, Ali Hummos*
(* - equal contribution, randomized order)

Code for Flexible task abstractions emerge in linear networks with fast and bounded units. For any questions about the code, contact Jan ([email protected]) or Alexandra ([email protected]).

Installation

To install the relevant libraries, run:

pip install -r requirements.txt

Code

For simulations, configurations of hyperparameters can be created using Config (in lcs/configs.py) and be run using run_script.py.

Figures

Notebooks to reproduce all simulations and figures in the manuscript (labeled by the respective figure) can be found in lcs/figure_notebooks/.

Citation

Please cite our paper if you use this code in your research.

@article{NTA2024,
  author = {Sandbrink, Kai J. and Bauer, Jan P. and Proca, Alexandra M. and Saxe, Andrew M. 
  and Summerfield, Christopher and Hummos, Ali},
  title = {Flexible task abstractions emerge in linear networks with fast and bounded units},
  publisher = {Advances in Neural Information Processing Systems},
  year = {2024}
}