Automatic Designs in Deep Neural Networks

Abstract

To train a Deep Neural Network (DNN) that performs well for a task, many design steps are taken including data designs, model designs and loss designs. Despite that remarkable progress has been made in all these domains of designing DNNs, the unexplored design space of each component is still vast. That brings the research field of developing automated techniques to lift some heavy work from human researchers when exploring the design space. The automated designs can help human researchers to make massive or challenging design choices and reduce the expertise required from human researchers.

Much effort has been made towards automated designs of DNNs, including synthetic data generation, automated data augmentation, neural architecture search and so on. Despite the huge effort, the automation of DNN designs is still far from complete. This thesis contributes in two ways: identifying new problems in the DNN design pipeline that can be solved automatically, and proposing new solutions to problems that have been explored by automated designs.

The first part of this thesis presents two problems that were usually solved with manual designs but can benefit from automated designs. To tackle the problem of inefficient computation due to using a static DNN architecture for different inputs, some manual efforts have been made to use different networks for different inputs as needed, such as cascade models. We propose an automated dynamic inference framework that can cut this manual effort and automatically choose different architectures for different inputs during inference. To tackle the problem of designing differentiable loss functions for non-differentiable performance metrics, researchers usually design the loss manually for each individual task. We propose an unified loss framework that reduces the amount of manual design of losses in different tasks.

The second part of this thesis discusses developing new techniques in domains where the automated design has been shown effective. In the synthetic data generation domain, we propose a novel method to automatically generate synthetic data for small-data object detection. The synthetic data generated can amend the limited annotated real data of the small-data object detection tasks, such as rare disease detection. In the architecture search domain, we propose an architecture search method customized for generative adversarial networks (GANs). GANs are commonly known unstable to train where we propose this new method that can stabilize the training of GANs in the architecture search process.