README.md

Geometric Ensemble for sample eXplanation (GEX)

Official code implementation of "GEX: A flexible method for approximating influence via Geometric Ensemble" (NeurIPS 2023)

How to use this repo?

Pull docker image for dependency

docker pull sungyubkim/jax:ntk-0.4.2

Run docker image

docker run -p 8080:8080/tcp -it --rm --gpus all \
--ipc=host -v $PWD:/root -w /root \
sungyubkim/jax:ntk-0.4.2

To run a single python file,

# to pre-train NN
python3 -m gex.pretrain.main \
    --dataset=mnist \
    --model=vgg \
    --corruption_ratio=0.1

# to estimate influence of pre-trained NN
python3 -m gex.noisy.main \
    --dataset=mnist \
    --model=vgg \
    --corruption_ratio=0.1 \
    --num_ens=8 \
    --ft_lr=0.05 \
    --ft_step=800 \
    --ft_lr_sched=cosine \
    --if_method=la_fge

To run multiple python files at once with ./gex/{task}/total.sh

bash gex/mnist/total.sh

Basically, results files (e.g., log, checkpoints, plots) will be saved in

./gex/{task}/result/{pretrain_hyperparameter_settings}/{posthoc_hyperparameter_settings}

Motivation: Identifying and Resolving Distributional Bias in Influence

Problem

As sample-wise gradient ($g_z$) follows stable distribution (e.g., Gaussian, Cauch, and Lévy), bilinear self-influence ($g_z M g_z$) follows unimodal distribution (e.g., $\chi^2$).

Key Idea

Influence Function can be interpreted as linearized sample-loss deivation (or more simply covariance) given parameters are sampled from Laplace Approximation.

$$ \mathcal{I}(z,z') = \mathbb{E}[ \Delta \ell^\mathrm{lin}(z, \psi) \cdot \Delta \ell^\mathrm{lin}(z', \psi)] = \mathrm{Cov}[\ell^\mathrm{lin}(z,\psi), \ell^\mathrm{lin}(z', \psi)]. $$

Solution

(1) Remove linearizations in sample-loss deviation and (2) Replace Laplace Approximation with Geometric Ensemble to mitigate the singularity of Hessian.

Supporting post-hoc methods

from gex.influence.estimate import compute_influence
# to compute influence kernel (N_tr, N_te) between train-test
influence_kernel = compute_influence(trainer, dataset_tr, dataset_te, dataset_opt , self_influence=False)
# to compute self-influence (N_tr) for train dataset
influence_kernel = compute_influence(trainer, dataset_tr, dataset_te, dataset_opt , self_influence=True)

• Random Projection (--if_method=randproj)
• TracIn Random Projection (--if_method=tracinrp)
• Arnoldi (--if_method=arnoldi)
• Laplace approximation with K-FAC (--if_method=la_kfac)
• Geometric Ensemble (--if_method=la_fge)

Acknowledgement

This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No.2019-0-00075, Artificial Intelligence Graduate School Program(KAIST))