From "A. S. Dogra and W. T. Redman, Optimizing Neural Networks via Koopman Operator Theory, Advances in Neural Information Processings Systems 33 (NeurIPS 2020)". This code implements Koopman training based on Node Koopman operators, which was what was used to obtain the results presented in the paper.
The provided function makes use of only Matlab.
An example data set of 25 NNs (~720 MB) trained using Ada delta (the data made to use Fig. 1) can be downloaded at https://drive.google.com/file/d/1GTz0osiZttg1VAd0WQaTvaasOHq8btjL/view?usp=sharing.
Having saved the example data to your computer, you can run HNN_master_example.m. This script calls the function NodeKoopmanTraining.m, which builds the Koopman operator and evolves the weights/biases forward. An error vs. weight/bias evolution plot (analogous to Fig. 1e) is produced. Various free parameters can be adjusted to get a feel for how Node Koopman training works. See the comments within the code and the paper for more details.
If you have any questions regarding the codebase or the associated NeurIPS paper, don't hesitate to email wredman@ucsb.edu or adogra@nyu.edu
