Black-box Predictions via Influence Functions

46927 Final Project

Authors: Ze Yang, Zhengyang Qi, Yundong Liu, Yuze Liu

Experiments

experiments.ipynb: Ridge on ForestFires dataset.
experiments2.ipynb: Binary Logistic Regression on Iris and MNIST.
experiments3.ipynb: Smoothed support vector on Iris.
experiments4.ipynb: Analysis of the components of loss influence function, using Marketing dataset.

Data

ForestFires, Iris, MNIST, Marketing: See /data/__init__.py

TODO

Report
Beamer Slides
More experiments; Images; Interpretations (Done)
Implement 2-Layers Perceptron (Done)
Implement (maybe) a simple CNN
LiSSA approximation (2) (Done)
Fancy Plots (Done)
Logistic Regression influence terms illustration (Done)
Implement Binary Logistic Regression (Done)
Implement Smoothed SVC (Done)
Implement Regularized Regression (Done)
Conjugate Gradient approximation (Done)
Improve Optimization Routine (Done)

Introduction

Our final project will be based on the paper “Understanding Black-Box Predictions via Influence Functions” by Pang Wei Koh and Percy Liang(1), which discusses how influence function(3) can trace the model’s prediction through the learning algorithm and the training data. For now, we decide to divide the projects into three parts. First of all, we will figure out the mathematical formula behind influence function. Secondly, we will study and implement the two techniques in efficiently calculating influence function discussed in the paper. Last but not least, we will apply the influence functions to other algorithms that we have learned in class and that are not discussed in the paper (e.g. Ridge Regression and Trees). Through this project, we intend to have a relatively thorough understanding of how influence functions can give us insights on training data.

Reference