Optimizing Training Data for Image Classifiers

Nikolaos Vasiloglou

Abstract:

In this paper, we propose a robust method for outlier removal to
improve the performance for image classification. Increasing the
size of training data does not necessarily raise prediction accuracy,
due to instances that may be poor representatives of their respective classes. Four separate experiments are tested to evaluate the
effectiveness of outlier removal for several classifiers. Embeddings
are generated from a pre-trained neural network, a fine-tuned network, as well as a Siamese network. Subsequently, outlier detection
is evaluated based on clustering quality and classifier performance
from a fully-connected feed-forward network, K-Nearest Neighbors
and gradient boosting model.