binarize-inverse-interpret

A tool for interpreting neural network image classifiers through activations inversion and analysis of binarized version of model output before classifier layer

Branches

• network_inversion
  Core branch with the full feature inversion pipeline for CNN and vision transformer models, including loss functions, logging, and visualization utilities.

• regnetx_bin_extractor_cpu
  CPU-adapted implementation for extracting binary codebooks from RegNetX models.

• resnet18_bin_extractor_cpu
  CPU-optimized extraction of binary codebooks from ResNet-18.

• regnetx_bin [archived]
  Analysis and extraction of binary features from the weights of RegNetX architectures.

• regnety_bin [archived]
  Binary feature analysis for RegNetY variants, which include Squeeze-and-Excitation blocks.

• resnet18_bin [archived]
  Binary feature analysis specifically for the ResNet-18 architecture.

Experimental networks sets

Binarization and SVD analysis:

• RegNetX_3_2
• ResNet18

Inversion:

• CNNs:
  • ResNet (versions larger than ResNet18 are recommended)
  • RegNetX
  • EfficientNetV2
  • ConvNexT
  • [will be improved soon] Wide_ResNet (reconstructs pictures with regular structures)
• visual transformers:
  • ViT
  • [will be improved soon] Swin (reconstructs random noise with with minor objects silhouettes)