Composition of simple computational tasks captures the inductive biases of animals in network models

This is the data is associated with a Nature Machine Inteligence manuscript: "Composition of simple computational tasks captures the inductive biases of animals in network models."

The repository contains partially processed data that was used to generate manuscript figures, as well as PyTorch .model files associated with trained networks. This data is associated with a github codebase: David Hocker. (2025). Savin-Lab-Code/kind_cl: Nature Machine Intelligence code (v1.0.0). Zenodo. https://doi.org/10.5281/zenodo.14907734

Data organization

rnndata: folder containing  RNN model and wait-time stats files for four types of pretraining curriculum methods. 

full_cl: performs all kindergarten tasks and behavioral shaping. (note: additional file type provided for this curriculum)

nok_cl: performs just behavioral shaping phases of pretraining

nok_nocl: performs no pretraining at all

pkind_mem: performs the memory pretraining task, followed by behavioral shaping

the model and stats files are for networks at the end of their training. the full_cl maniuplation contains additional data, as it was the main form of curriculum learning studied in the manuscript. python-compatible forms of the behavior over training are provided (rnn_#_allbeh.json files). RNN 33 also has additional files. it contains model files over the full course of learning, as well as a sample set of simulated data at the end of training (full_cl/33/rnn_curric_33_block_10_1k.json)

processed4plots: folder containing intermediate-level processed data for visualizing results presented in the manuscript. These files are called from visualization scripts in the corresponding github repo (https://github.com/Savin-Lab-Code/kind_cl/)

behdat_rnn_full_cl_end/: subfolder containing matlab files for behavior of each network from full_cl at end of training

behdat_rnn_full_cl_overtraining/: subfolder containing matlab files that aggregate behavior of each network from full_cl across training

behdat*: aggregated features of RNN behavior over training, for a given curriculum type

flowfields*: .mat matlab files that provide 2D flow fields for RNNs 33 and 23

ke_overtraining*: aggregated dynamical systems features measures (number of slow features, dimensionality) over training, for a given curriculum type

ke_results*: distributions of slow features and dimsionality for 3 main curricula presented in Figure 6E

kemin_rnn_curric*: converted .mat version of results of kinetic energy minimization. provides location of slow points in low-dimensional PC space

rrate*: aggregated reward rates of RNNs over training, for a given curriculum type