Learning Hierarchical Generative Models: Theory and Applications

近年来，计算能力和可从网络、摄像机、高通量基因组测序技术和各种实验室测量获得的数据量大幅增加。构建能够自动从这些数据中发现有意义的表示的智能系统，应该会带来新的科学发现。例如，它们可以帮助神经科学家分析高维fMRI大脑成像数据，或改进亚马逊等公司的产品推荐系统。 我的主要科学兴趣是了解在大量数据中发现结构所需的计算原理。我提出的研究集中在开发一个新的框架，用于表示和学习支持多层次推理的大规模深度生成模型。这一类新的分层概率模型为定义高维数据上灵活的概率分布提供了强大的工具，并允许我们构建丰富的概率模型，可以自动从大量高维数据中发现语义规则性、结构化关系或不变性。 目前，许多现有的机器学习系统，如支持向量机，从根本上限制了它们从高维数据中学习复杂结构关系的能力。学习系统无法科普未经专门训练的新任务。我的研究旨在开发多功能的概率模型，能够从数据中提取高阶知识，并成功地将获得的知识转移到学习新任务中。这些模型对许多研究领域产生重大影响，包括计算生物学，神经科学，医学诊断，数据挖掘和机器人技术。

In recent years, there has been a massive increase in computational power and the amount of data available from the web, video cameras, high-throughput genomic sequencing technologies, and various laboratory measurements. Building intelligent systems that can automatically discovery meaningful representations from such data, should lead to new scientific discoveries. For example, they can help neuroscientists analyze high-dimensional fMRI brain imaging data, or improve product recommendation systems of companies like Amazon. My main scientific interest is to understand the computational principles required for discovering structure in large amounts of data. My proposed research concentrates on developing a novel framework for representing and learning large-scale deep generative models that support inferences at multiple levels. This new class of hierarchical probabilistic models provides a powerful tool for defining flexible probability distributions over high-dimensional data, and allows us to build rich probabilistic models that can automatically discover semantic regularities, structured relations, or invariances from large volumes of high-dimensional data. Many existing machine learning systems today, such as support vector machines, are fundamentally limited in their ability to learn complex structural relations from high-dimensional data. Learning systems cannot cope with novel tasks for which they have not been specifically trained. My research aims to develop probabilistic models that are multi-functional, capable of extracting higher-order knowledge from data, and successfully transfer acquired knowledge to learning new tasks. These models hold great promise for making a big impact on many research areas, including computational biology, neuroscience, medical diagnosis, data mining, and robotics.