Description
Small CNN Experiment Data Repository
DataBase POC: andrew.engel@pnnl.gov
This repository contains the data products created for the https://arxiv.org/pdf/2310.18612,
"Efficient kernel surrogates for neural network-based regression", submitted to PNAS. As
outlined in section 6.3, a small CNN model was trained to classify CIFAR10, MNIST, and FMNIST.
After every epoch the model weights were saved, the conjugate kernel was computed and saved,
the neural tangent kernel was computed and saved, and a linear probing model where the final
weight vector of the original NN model was allowed to continue training for 5 epochs was saved.
This repository contains dataproducts and logs allowing one to train kernel surrogate models
using the pre-computed CK and NTK kernels, as well as verify the weights of our NN model and linear
probes. A complete description of the data can be found in the attatched data directory,
('data_dict.txt') accompanying the repository. All contained files are binary array data, either
marked with the .npy or .pt file extension, to be opened with either np.load or torch.load commands.