Skip to content

README.md

Latest commit

 

History

History
49 lines (40 loc) · 2.33 KB

README.md

File metadata and controls

49 lines (40 loc) · 2.33 KB

learning-2-learn

L2L Sensorimotor Association problems (Nat. Neuro. 2023)

Code dependencies

  • Python (model)
    • Tensorflow
    • Numpy
    • Scipy
  • Matlab (analysis)

Model Training

Model can be trained via a call to the function train() in train.py

Call with default parameters used in the paper looks like:

train(seed          = 0,
      batchSize     = 1,
      l2            = 0.0005,
      l2_wR          = 0.001,
      l2_wI          = 0.0001,
      l2_wO          = 0.1,
      learningRateInit = 0.0001,
      svBnd = 10.0,
      rType = runType.Full)

Here parameter(s):

  • starting with $l2$ are regularization hyper-parameters
  • $svBnd$ is the number of recurrent weight matrix singular values to regularize
  • $learningRateInit$ is the learning rate
  • $batchSize$ indicates 1 trial is learned at a time - keep this unchanged
  • $rType$ indicates:
    • runType.Full : Train 1001 problems, one after the other
    • runType.ControlManifPert : Control condition for manifold perturbation; train 50 problems to uncover decision and stimulus subspaces; train on another 50 problems with frozen output weights.
    • runType.SSManifPert : S → S manifold perturbation; after uncovering decision and stimulus subspaces; train on another 50 problems with independent S → S perturbations
    • runType.DSManifPert : D → S manifold perturbation; after uncovering decision and stimulus subspaces; train on another 50 problems with independent D → S perturbations

Raw output files (trained weights - [file prefix saved] and trials to convergence - [file prefix conv]) are saved to the data folder. This folder is currently empty, but raw pre-trained model data in zipped format may be downloaded from here.

Model Analysis

Model analysis code that replicates main figures in paper are located in the analysis folder. This code saves processed data to the results folder.

Each matlab script in the analysis folder replicates 1 or 2 main figures. Each script that relies on time-consuming processing starts with the line:

useSaved = true;

To re-run the processing on the raw files, set this flag to false. To rapidly create the figures from the saved processed data in the results folder, leave this flag set to true.