Efficient Lifelong Learning with Structured and Graphical Representation: Benchmarks, Methodologies and Applications

高效持續學習的結構化和圖模型表示:基準測試、算法和應用

Student thesis: Doctoral Thesis

View graph of relations

Author(s)

Related Research Unit(s)

Detail(s)

Awarding Institution
Supervisors/Advisors
Award date2 Sept 2024

Abstract

Lifelong learning is an intrinsic ability in human cognitive learning processes. However, for deep learning systems, achieving lifelong learning is challenging. Ideally, lifelong learning agents can continuously acquire new knowledge while retaining old knowledge and integrating it to enhance learning new information in an adaptive and compact way. This difficulty for deep lifelong learning primarily arises due to catastrophic forgetting and the limitations in transferability inherent in current learning frameworks. Moreover, it is highly impractical and inefficient to train separate models for each task, data distribution, and scenario, making such an approach unusable. Therefore, it is crucial to design algorithms that enable deep learning models to possess lifelong learning capabilities similar to humans, while also being efficient and practical. In this thesis, we primarily aim to address the following questions: (1) How can we define time-evolving tasks in real-world scenarios, such as robotic vision, within the context of lifelong learning? How can we establish benchmarks and evaluation protocols for lifelong learning from task-specific and practical perspectives, including various semantics of task transfer and forgetting, difficulty levels, similarity, and computational efficiency complexity? (2) Based on these practical perspectives, how can we design efficient and interpretable methods, inspired by the human learning process and neuroscience, and how can these methods be implemented in specific applications?

This thesis mainly includes two main parts. In the first part, we analyze the existing literature on the definition, challenges, and solutions of lifelong learning. We will then reassess and benchmark these definitions and evaluation standards from a practical application perspective, providing a benchmark in the context of robotic vision, characterized by dynamic environments, multiple objects, and high transferability.

In Part II, we aim to propose efficient lifelong learning algorithms by disentangling the lifelong learning process into two aspects: model and data. From the model perspective, we will investigate the structural and sparsity topologies, and inductive biases that can achieve efficient lifelong learning. We analyze deep learning models from the viewpoint of graph topology. Inspired by structural characteristics found in human neural networks, such as small-world structures and scale-free networks, we identify that similar implicit representations exist in deep learning models, especially sparse models. We will leverage these discoveries to explicitly enhance these topological structures in the networks and dynamically evolve the networks based on these topological dynamics in evolving tasks. The framework aims to enable deep learning models to perform well across various tasks, including different data fidelity and different modern deep learning tasks (such as single-task learning, network compression, network pruning, lifelong learning, and few-shot learning), while improving computational efficiency thanks to the sparsity and evolution efficiency of these topological structures. From the data perspective, we explore the memory mechanisms in lifelong learning akin to human memory, identifying the most efficient and representative memory data that is best suited for network optimization. We provide a comprehensive evaluation and an effective mixed memory mechanism solution.

In summary, this thesis aims to bridge the gap between algorithmic research and practical applications in lifelong learning for deep learning. By providing benchmarks and evaluation protocols, proposing novel algorithms that leverage structural and data perspectives, and demonstrating their effectiveness in real-world applications, the thesis seeks to advance the field of lifelong learning. The proposed frameworks not only enhance performance and computational efficiency but also offer a practical and scalable approach to implementing lifelong learning in diverse and dynamic environments. Additionally, these practical benchmarks and efficient lifelong learning methods may provide potential advancements in the lifelong learning of new paradigms, such as large language models, vision-language models, and other new advances in foundational models.