In ansatz_verification, we present code for verifying that AGOP is highly correlated with Neural Feature Matrices (NFMs) for transformers, CNNs, RNNs, and MLPs. In each directory, we provide a README outlining the requirements needed to run corresponding code.
In feature_visualizations, we provide code for visualizing features captured by AGOP of CNNs (an example is provided for VGG-19) and GPT2-architecture language models (an example is provided for a model trained on the TinyStories dataset).
In kernel_machine_agop, we provide code for computing AGOP of trained kernel machines using Laplace kernel and Convolutional Neural Tangent Kernel (CNTK).