What problem does it solve? Researchers studying feature learning in convolutional neural networks need reproducible tooling to verify the convolutional neural feature ansatz (CNFA), compute expected gradient outer products (EGOP), and train deep convolutional recursive feature machines, which otherwise requires writing patch-based Jacobian and kernel analysis code from scratch. ## Core Features & Use Cases - CNFA Verification: Compute patch-level Jacobians and EGOP matrices on pretrained ResNet and VGG networks to test whether learned kernels match the neural feature ansatz. - Deep ConvRFM Training: Iteratively train convolutional recursive feature machines with gradient-based kernel updates for binary and multiclass image classification. - Kernel Visualization: Compute and visualize VGG kernel eigenvectors and propagate them through network layers for interpretability analysis. - Use Case: A researcher investigating feature learning on CIFAR or ImageNet can patchify inputs, compute EGOP over a trained VGG network, and compare the resulting feature matrix against the learned convolutional kernels. ## Quick Start Clone the convrfm repository, install PyTorch and its dependencies, then run the EGOP and patchify demo script to compute an expected gradient outer product on synthetic image data.
