radaimi / Understanding-Models-Through-The-Training-Data

Understand predictions of black box models by studying the influence of the training data

"Understanding Models Through The Training Data"

In this repository, we replicated some of the experiments discussed in Koh et al's paper titled "Understanding Black-box Predictions via Influence Functions". The original code can be found here: http://bit.ly/gt-influence.

We focused on two use cases of influence functions discussed in the paper:

1. Understanding model behavior
2. Debugging domain mismatch

For each task, we replicated the results presented in the paper and implemented additional experiments using other datasets or other models. For a more detailed description of our work and our findings, kindly refer to the report.

The datasets and model weights used for the first application "Understanding model behavior" can be downloaded:

1. Inception V3 weights
2. Inception ResNet V2 weights
3. DogFish Dataset (From Koh's replication package of the paper)