Finding Influential Examples in Deep Learning Models

Abstract

Machine learning models are powerful, but not without errors and the complexity of large models makes it hard for a human to intuitively understand the cause of the error. This thesis approaches the task of explaining predictions made by deep learning models by studying the importance of specific examples in the training data, referred to as influence. In practice, the embedding representation of the training data, defined as the output from an arbitrary layer in the model, is compared to the influence on a prediction. Two models are investigated; a Logistic Regression model and a Convolutional Neural Network. The aim of this thesis is thus to identify influential examples in deep learning models in a computationally efficient way, by studying the relation between the representation of the data in a network and its influence. The main results include comparisons between various metrics of distance in the embedding representation of the images to their influence. Similar examples are shown to be clustered close together, training examples close to a test example exhibited high influence for a correctly classified test example. Training examples far away from its class centroid in the embedding space also show high influence.