Knowledge Transfer in Artificial Intelligence Systems

不断获取和转移知识是人类智能的一个关键特征，这使我们能够迅速适应新的任务和环境，因为我们不必从头开始学习。通过知识转移，我们不断发展各种复杂的能力来处理不同的任务和问题。虽然这是人类智能的固有能力，但如何使人工智能系统适当地转移知识在很大程度上仍然是一个未解决的问题。另一方面，最近机器学习技术的成功在很大程度上依赖于高效算法的发展，以及大数据集的可用性。然而，在许多现实世界的应用中（例如，医学，神经科学），收集数据要么昂贵要么耗时，这使得“大数据”无法构建可靠的机器学习模型。知识转移在这些应用中是至关重要的，因为它可以帮助在先验知识的帮助下从小数据中学习。为此，我提出的研究计划将集中于开发一种通用的持续知识维护和转移（CKMT）系统的综合方法，该系统可以通过知识转移快速学习新任务并适应不断变化的环境，而只需很少的训练数据。更具体地说，拟议的研究是围绕三个主要目标组织的：1。创建机器学习任务之间安全转移的新机制；2. 任务的非平稳环境建模；3. 创建跨不同机器学习范式的知识转移框架。我们还将运用我们的方法解决有实际影响的问题。特别是将其应用于各种应用场景的脑信号分析，以及电力系统运行和智能电网能源管理。本项目将解决人工智能中的一些基本问题，包括迁移学习、多任务学习、元学习和终身学习。它将为知识转移提供新的理论见解和理由，并激发新的知识转移算法和模型。为了实现每一个目标，我将培养一名博士和一名硕士学生，为他们提供人工智能和机器学习方面非常抢手和有价值的技能，这些技能可以转移到许多学科，比如生物医学工程、神经科学、电力/能源管理，以及其他学科，比如金融/银行/保险。

Continually acquiring and transferring knowledge is a key characteristic of human intelligence, which enables us rapidly to adapt to new tasks and environments as we do not have to learn from scratch. By knowledge transfer, we continually develop a wide variety of complex abilities to deal with diverse tasks and problems. While it is an inherent ability of human intelligence, how to enable artificial intelligence systems to appropriately transfer knowledge remains a largely unsolved problem. On the other hand, the recent success of machine learning techniques largely relies on, in addition to the development of efficient algorithms, the availability of large data sets. However, in many real-world applications (e.g., medicine, neuroscience), collecting data is either expensive or time consuming, which makes "big data" unavailable to build reliable machine learning models. Knowledge transfer would be critical in these applications, as it can help to learn with small data with the help of prior knowledge. To this end, my proposed research program will be focused on the development of a comprehensive approach to general-purpose continual knowledge maintain and transfer (CKMT) systems, which can rapidly learn new tasks and adapt to changing environments, with little training data, by knowledge transfer. More specifically, the research proposed is organized around three main objectives: 1. Creating a new mechanism for safe transfer between machine learning tasks; 2. Modeling the non-stationary environments of tasks; 3. Creating a framework for knowledge transfer across different machine learning paradigms. We will also apply our methods to solve problems with practical impact. In particular, we will apply them to brain signals analysis in various application scenarios, as well as power system operation and smart grid energy management. This project will address some fundamental issues in Artificial Intelligence, including transfer learning, multitask learning, meta-learning and lifelong learning. It will provide novel theoretical insights and justifications of knowledge transfer, and motivate new knowledge transfer algorithms and models. With each objective, I will train one PhD and one MSc student, providing them with highly sought-after and valuable skills in Artificial Intelligence and Machine Learning transferable to many subjects, such as biomedical engineering, neuroscience, power/energy management, and others like finance/banking/insurance, for example.