Better Generalization with Less Human Annotation Using Meta-Learning and Self-Supervised Learning for Image Analysis

Deep neural networks require a large amount of annotated training data to generalize well. Unfortunately, such large training data is difficult to obtain in medical or geophysical domains due to many reasons, including the high annotation cost, privacy concerns, and physical constraints. To address the data efficiency issue in training the deep learning model, we proposed the solutions in two directions: meta-learning and self-supervised learning. Meta-learning tries to generate a robust model that can learn to quickly adapt to new tasks with minimal labeled samples. It is also called “learning to learn”, which acquires fast adaptation capability over a collection of related tasks and uses it to improve its future learning performance. Self-supervised learning, on the other hand, tries to leverage all useful information from the unlabeled training data itself. It can improve feature representation by solving a pretext task, or directly solving the unsupervised target task which shares the same task structure as the self-supervised task. In the first part of this work, we introduce the meta-learning setting, develop a meta-learning framework AGILE+ to deliver the efficient rat brain cell classifier, and study the first arrival picking problem and solve the domain shift problem with less human interaction. In the second part, we discuss different self-supervised learning settings, introduce the 3D self-supervised image patch reconstruction task to significantly improve the incidental lung nodule classification accuracy for the data-hungry 3D Vision Transformer (3D-ViT) model, and propose the self-supervised learning model Blind-Trace Network (BTN) for the application in the unsupervised seismic interpolation task.