Deep Learning with Applications in Pattern Recognition and Image Analysis

机器学习的目标是设计算法和统计模型，通过从示例中学习而不是遵循特定的编程指令来执行“智能任务”，例如对象和语音识别、生物识别、计算机视觉、自动医疗诊断和数据挖掘。随着社会中大数据集的激增和大数据革命的开始，机器学习已经成为当代信息技术的关键组成部分，因为它促进了大量数据的高效搜索和挖掘，以及数据的智能处理。拟议的研究将侧重于设计和分析新的机器学习算法，以应对通常难以大量获取的多维、非结构化和无注释数据（例如医疗数据）所带来的挑战。新的和实质性的研究挑战出现在算法的分析和设计中，这些算法学习数据分布适应数据非平稳性，并结合对内存的实际约束，导致过度参数化的算法复杂性，训练和计算速度的限制以及全局优化的挑战。

The goal of machine learning is to design algorithms and statistical models that perform “intelligent tasks” such as object and speech recognition, biometrics, computer vision, automated medical diagnosis and data mining, by learning from examples rather than following specific programming instructions. With the proliferation of big data sets in our society and onset of the big data revolution, machine learning has become a key component in contemporary information technologies as it facilitates efficient search and mining of large volumes of data, as well as intelligent processing of the data. The proposed research will focus on designing and analyzing novel machine learning algorithms that meet the challenges posed by multidimensional, unstructured and not annotated data often difficult to acquire in large quantities (e.g., medical data).. New and substantial research challenges arise in the analysis and design of algorithms that learn data distributions adapt to data nonstationarity and incorporate practical constraints on memory, complexity of algorithms leading to overparametrization, limits on training and computational speed and challenges in global optimization. The research objectives are divided into two main themes: (1) Application of computational learning theory and complexity regularization to analysis of deep convolutional and multilayer neural networks, deep random forest classifiers in order to analyze their convergence, rate of convergence and learning speed. Design of new efficient and accurate kernel classifiers with superkernels; (2) Application of principal curves and manifolds in medical imaging for robust and accurate segmentation of cytological and histopathological slides, counting and classifying white and red blood cells and for 3D segmentation of tumors in ultrasound and in tomosynthesis breast images. Automatic feature extraction from medical images by the deep convolutional neural networks and combining them with powerful classifiers such as SVM and random forest. The training component of the proposed research will provide 2 M.Sc. and 3 Ph.D. students each year with stimulating research challenges and immerse them in important current topics in machine learning. The research is expected to provide a deeper understanding of the fundamental principles of learning in deep networks as well as its practical aspects, such as fast and easy implementable algorithms with good performance.