高效贝叶斯最大间隔特征空间变换方法研究

With the distinct characteristics such as generative modeling, automatic parameter estimation, strong interpretability and good robustness, Bayesian learning has made remarkable achievements in many application fields. But its generally inefficient inference and the inconvenience of combining with the popular maximum margin learning principle in machine learning field, has been the obstacles for its further success in industry. Especially with the situations that the accumulated data in every walk of life is more and more, the dimension is higher and higher, the flow speed is faster and faster, traditional Bayesian learning is faced with great challenge. To address these problems, this project studies the following efficient Bayesian maximum margin learning algorithms:..First, to efficiently process high-dimensional, multi-view data, we propose a Bayesian maximum margin multiple kernel nonlinear feature transformation algorithm based on pseudo likelihood, data augmentation and Riemannian manifold Hamiltonian monte carlo;..Second, to improve the flexibility and efficiency of processing large-scale training data set, we propose a parallel infinite mixture of local maximum margin feature transformation algorithm based on the Dirichlet process;..Third, for efficient and accurate modeling of streaming data, we propose an online Bayesian maximum margin linear subspace learning algorithm and an online Bayesian maximum margin multiple kernel nonlinear feature transformation algorithm, based on the posterior regularization technique.

贝叶斯学习以其生成式建模、参数自动推断、模型自适应性与鲁棒性强等鲜明特点在许多应用领域已取得了令人瞩目的成就。但其求解普遍低效的问题以及难以与机器学习领域流行的最大间隔学习准则相结合等一直是阻碍其进一步被工业界接受的重要因素。特别是随着各行各业积累的数据越来越多，维度越来越高，流动速度越来越快，传统贝叶斯学习面临着极大挑战。为加以应对，本项目研究以下高效贝叶斯最大间隔学习算法：.第一，从高效精准地处理高维、多视图数据出发，提出基于伪似然与数据扩充以及黎曼流形哈密尔顿蒙特卡罗的贝叶斯最大间隔式多核非线性特征空间变换算法；.第二，为提高处理大规模训练数据集的灵活性和效率，提出基于狄利克雷过程的无穷混合局部最大间隔特征空间变换算法以及相应的并行后验推断算法；.第三，为高效精准地对流式数据建模，提出基于后验正则化的在线贝叶斯最大间隔线性子空间学习算法和在线贝叶斯最大间隔多核非线性特征空间变换算法。

贝叶斯学习以其生成式建模、参数自动推断、模型自适应性与鲁棒性强等鲜明特点在许多应用领域已取得了令人瞩目的成就。本课题基于贝叶斯统计机器学习方法和概率生成模型开展一系列高效、 鲁棒、 自适应、 判别性强的特征空间变换算法研究及其应用，在贝叶斯最大间隔学习、深度生成模型、多视图学习、神经信息解码、多模态学习、半监督学习、在线学习等重要领域总计取得了 10 项理论和应用成果。这些成果在一定程度上解决了训练样本不充足、噪声大、维度高、流速快等挑战性难题，较明显地提升了相关应用的实际效果，为人工智能技术的进一步推广应用做出了贡献。同时，部分成果得到了国内外关注，或被推选为会议最佳论文，或被多次报道评论。