Deep Learning With Very Few Training Examples

Zusammenfassung

This dissertation addresses the problem of training deep learning models with very few training examples. While deep learning has achieved remarkable success across a wide range of domains, deep learning models typically have a vast number of parameters that need to be optimized, and large amounts of labeled data are required for training. However, the collection and annotation of thousands or millions of examples is extremely time-consuming and expensive. This is a significant limitation of deep learning methods in many real-world applications. In the first part, a novel object detection method is proposed for recognizing new categories with very few training examples by combining the advantages of convolutional neural networks and random forests. Subsequently, a new method called Neural Random Forest Imitation (NRFI) is presented, designed to implicitly transform random forests into neural networks. The experiments demonstrate that NRFI is scalable to complex classifiers and generates very small networks. Finally, two novel generative methods, ChimeraMix and HydraMix, are presented for small data image classification, which learn the generation of new image compositions by combi...