Abstract representations emerge naturally in neural networks trained to perform multiple tasks

These are the simulation data that underly some of the figures in this paper: https://doi.org/10.1101/2021.10.20.465187 In particular, these data save the arguments, generalization metrics, and some other details from the simulations showing how generalization depends on the number of tasks. In most cases, they do not save the underlying model or samples from the underlying model -- though the code for training such models is provided in the github repository linked below. More information about these data and code for reading them and replotting the figures in the paper can be found here: https://github.com/wj2/disentangled