influence-function-reproduction

This repository contains a PyTorch implementation reproducing Figure 2 (Linear Approximation vs. Actual Loss Change) from the paper Understanding Black-box Predictions via Influence Functions (Koh & Liang, 2017).

Implementation Details

• Method: Primitive PyTorch implementation (Calculated gradients and Hessian-Vector Products using autograd).
• Optimizer: L-BFGS .
• Hessian Inverse: LiSSA .
• Dataset: Binary MNIST.

Files

• figure2_linear_approx.py: Main script to run the experiment and generate the plot.
• influence_functions.py: Core implementation of gradient and HVP calculations.

Results

The reproduction demonstrates a near-perfect linear correlation between the predicted influence and the actual loss change.

How to Run

1. Install dependencies:

   pip install torch torchvision numpy matplotlib tqdm

2. Run the reproduction script:

   python figure2_linear_approx.py