jax_influence_function

A Google Colab notebook that trains a neural net in JAX on the Atis Intent dataset and uses the Understanding Black-box Predictions via Influence Functions paper to generate influence scores for each data point in the training set for a specific test point. The notebook ends by retraining the net while removing the least helpful points, to demonstrate that performance on the test point has increased.