Automated Machine Learning and Meta-Learning for Multimedia

Table of Contents

1. Frontmatter

2. Part I

   1. Frontmatter

   2. Chapter 1. Automated Machine Learning

      Wenwu Zhu, Xin Wang

      Abstract

      The last decade has witnessed a surge of machine learning (e.g., deep learning) research and applications in many real-world scenarios, such as computer vision, language processing and data mining. Most machine learning methods have a plethora of design choices that need to be made beforehand, and their performance is shown to be very sensitive to these choices. Furthermore, the desirable choices of algorithm design often vary over different tasks and hence the algorithm configuration requires intensive expertise, which becomes a substantial hurdle for new users and further restricts the applicability and feasibility of modern machine learning methods in a wider range of public fields.

   3. Chapter 2. Meta-Learning

      Wenwu Zhu, Xin Wang

      Abstract

      Last decade has witnessed a prosperous development for supervised learning, i.e., learning tasks with given labels for model training. Supervised learning usually depends on large labeled datasets and trains a huge model with a large number of parameters from scratch. Thus, the requirement for data and computing resources is relatively high. However, there are many applications where data is difficult or expensive to collect, or computing resources are limited. Since the lack of training data, supervised learning is not suitable for these tasks and shows bad performances.

3. Part II

   1. Frontmatter

   2. Chapter 3. Automated Machine Learning for Multimedia

      Wenwu Zhu, Xin Wang

      Abstract

      The term Multimedia has been taking on different meanings from its first advent in 1960s until today’s common usage which refers multimedia to “an electronically delivered combination of media including videos, still images, audios, and texts in such a way that can be accessed interactively”.

   3. Chapter 4. Meta-Learning for Multimedia

      Wenwu Zhu, Xin Wang

      Abstract

      Recently, meta-learning has seen a drastic rise of interest especially in deep learning due to its ability to tackle conventional challenges in deep learning, e.g., data and computational bottleneck, generalization ability, etc. Meta-learning is most often applied to alleviate the data scarcity problem in the few-shot learning paradigm, deriving prior knowledge across multiple learning tasks for rapid adaption to new tasks with limited amount of training data. In multimedia context, meta-learning has great potential to be applied to various practical scenes to improve data efficiency and generalization ability, while more exploration is still needed for future researchers. In this chapter, we discuss meta-learning for two application categories: i) multimedia search and recommendation, including cold-start recommendation, recommender algorithm selection, and incremental product search; ii) vision and language, including classification, detection, image captioning, tracking, and visual question answering problems.

   4. Chapter 5. Future Research Directions

      Wenwu Zhu, Xin Wang

      Abstract

      In this chapter, we provide the readers with several potential future research directions on AutoML (hyper-parameter optimization and neural architecture search) and meta-learning. We would like to point out that the future works deserving further investigations discussed in this chapter can also be applicable to various multimedia applications as well as the learning tasks of different modalities (i.e., texts, audios, images, and videos), therefore we do not present separate discussions here.

4. Backmatter