Iterative Label Improvement: Robust Training by Confidence Based Filtering and Dataset Partitioning

State-of-the-art, high capacity deep neural networks not only require large\namounts of labelled training data, they are also highly susceptible to label\nerrors in this data, typically resulting in large efforts and costs and\ntherefore limiting the applicability of deep learning. To alleviate this issue,\nwe propose a novel meta training and labelling scheme that is able to use\ninexpensive unlabelled data by taking advantage of the generalization power of\ndeep neural networks. We show experimentally that by solely relying on one\nnetwork architecture and our proposed scheme of iterative training and\nprediction steps, both label quality and resulting model accuracy can be\nimproved significantly. Our method achieves state-of-the-art results, while\nbeing architecture agnostic and therefore broadly applicable. Compared to other\nmethods dealing with erroneous labels, our approach does neither require\nanother network to be trained, nor does it necessarily need an additional,\nhighly accurate reference label set. Instead of removing samples from a\nlabelled set, our technique uses additional sensor data without the need for\nmanual labelling. Furthermore, our approach can be used for semi-supervised\nlearning.\n