Interpreting Neural Networks With Nearest Neighbors
This work addresses the issue of interpretability for users of neural networks, offering an incremental improvement over existing local interpretation methods.
The paper tackled the problem of unreliable feature importance in neural network interpretation by using Deep k-Nearest Neighbors to provide a robust uncertainty metric, resulting in interpretations that better align with human perception than baseline methods.
Local model interpretation methods explain individual predictions by assigning an importance value to each input feature. This value is often determined by measuring the change in confidence when a feature is removed. However, the confidence of neural networks is not a robust measure of model uncertainty. This issue makes reliably judging the importance of the input features difficult. We address this by changing the test-time behavior of neural networks using Deep k-Nearest Neighbors. Without harming text classification accuracy, this algorithm provides a more robust uncertainty metric which we use to generate feature importance values. The resulting interpretations better align with human perception than baseline methods. Finally, we use our interpretation method to analyze model predictions on dataset annotation artifacts.