Add MA-GIG explainer

[leekwoon]

Adds MA-GIG (Manifold-Aligned Guided Integrated Gradients, ICML 2026) as a PnPXAI explainer, with a worked tutorial.

Paper: https://arxiv.org/abs/2605.02167  ·  Reference code: https://github.com/leekwoon/ma-gig

What it does

Guided IG picks the features with the smallest gradients at each step and moves them toward the input. That update is axis-aligned in pixel space, and the tangent space of natural images is not, so the path drifts off the data manifold and the error accumulates along it. MA-GIG runs the same greedy selection inside the latent space of a pretrained VAE: an axis-aligned step in latent space becomes a column of the decoder Jacobian in pixel space, i.e. a tangent vector, so the decoded path stays close to the manifold and gradients are read off plausible images.

Contents

File
pnpxai/explainers/magig.py	the explainer, self-contained (no dependency on the paper's repo)
tutorials/MAGIG_example_all_explainers.md	worked example on Oxford-IIIT Pet with the paper's fine-tuned ResNet-18
tests/explainers/test_explainers.py	10 tests using a stub autoencoder, so they run in ~5 s with no download
docs/api/explainer/magig.md, mkdocs.yml, README.md	registration

Verification

Attributions are bit-identical to the reference implementation at the paper's configuration (n_steps=200, fraction=0.05, slerp, black baseline) — max|diff| = 0.0, Pearson r = 1.0 on ResNet-18 and GoogLeNet.

On the tutorial's 50-image benchmark, DiffID is 0.476 for MA-GIG against 0.347 for IntegratedGradients started from the same baseline: a gap of +0.129, against the +0.085 the paper reports for this dataset and classifier.

Notes for review

Not in AVAILABLE_EXPLAINERS. Constructing MAGIG downloads an autoencoder, and one explanation runs 200 decoder passes, so including it in auto-recommendation would make every AutoExplanation heavy. The tutorial adds it explicitly with manager.add_explainer, as the LEAR tutorial does. Happy to change this if you would rather it be recommended.

No new checkpoints hosted. The paper uses the Stable Diffusion 2.1 autoencoder; stabilityai/sd-vae-ft-mse publishes weights identical to it (verified: maximum parameter difference 0.0), so it loads from there — the same pattern gfgp.py uses for its generative model. The tutorial's classifier comes from the paper's own repository.

cudnn.deterministic matters for this method. The greedy step selects latent dimensions by a low quantile of gradient magnitude, where the distribution is dense — roughly 130 of 4096 dimensions sit within 1e-7 of the threshold — so a last-bit difference flips which dimensions move and the rest of the path diverges. Under PyTorch's default non-deterministic cuDNN setting, two identical calls return visibly different maps; with torch.backends.cudnn.deterministic = True they are bit-identical. The constructor warns when the flag is unset on CUDA, and the tutorial sets it.