Abstract

This thesis tries to investigate the impact of label noise on the performance of binarized neural networks (BNNs). Our experiments are evaluated on MNIST and CIFAR-10 where we test sev- eral architectures including multi-layer perceptrons (MLPs) and convolutional neural networks (CNNs) of varying depths. We show that BNNs are able to handle large amounts of label noise. In one experiment we verified that when diluting every clean CIFAR-10 sample with 10 noisy samples, binarized CNNs can still attain over 50% accuracy. This behaviour also holds for MLPs on MNIST where we saw a prediction accuracy of about 65% even when the amount of noisy labels per clean label was as high as 100.