Efficient and Reproducible Image Annotation for Supervised Deep Learning with Small Data

点击翻译按钮获取中文摘要

Machine learning is currently one of the hottest and most quickly evolving technologies. Everyone wants to use "deep learning" to develop systems for everything: face recognition, identity tracking, quality control in manufacturing, tracking objects in videos, diagnosing disease, precision agriculture, etc. Deep learning works best when there are very large numbers of images where the "correct answer" is known from which the machine learning algorithm can learn. It is relatively easy for people to tag an image of a car, a boat, or a plane, or even different types of cars with an appropriate label. It is much more challenging to create large datasets to train systems for diagnosing cancer, or counting the number of flowers on a plant. Labeling medical images requires highly experienced experts. When marking all of the flowers on a plant, it is easy to miss one, or for a flower to be partially occluded by a leaf or another flower, or to be imprecise in specifying a flower's exact location. Moreover, in order to obtain enough annotated images to learn from, typically multiple annotators label a dataset but each image is annotated by only one annotator. In such circumstances we have no way of determining whether different annotators are annotating consistently, that is, we can't quantify the inter-annotator agreement. Annotator disagreement arises from biases in their annotations, and degrades the quality of the training dataset. Inter-rater agreement is affected by the difficulty of the annotation task and the nature of the instructions given to annotators. If we can find ways of obtaining more consistent annotations across multiple annotators, training dataset quality will improve, and hence the performance of the learned system also improves. The proposed research program will study how to obtain better quality annotations from annotators with higher inter- and intra-annotator agreement. We will create augmented annotation tools that provide problem-specific semi-automation to assist annotators and quantify the resulting benefits to annotator agreement and trained system performance. We will quantify the relationship between annotator agreement and model performance. We will explore the degree to which contextual factors such as annotation type, instructions given, pressure, and distractions can influence annotator agreement and develop best practices for mitigating their effects. By studying factors that influence annotator agreement and the performance of the systems that are learned from annotated datasets, we will be able to develop new standardized methodologies for training "deep learning" models with limited data. This will allow better prediction of the optimal amount of resources to invest in annotation, reduce the reliance on trial-and-error methods to obtain the best trained system performance, and make successful machine learning less reliant on deep technical expertise.