CIF: Small: Active data screening for efficient feature learning

传感和数据采集技术的进步使得产生大量必须存储、交流、处理和理解的数据变得容易。这些数据的质量可能是可变的，数据的性质可能随着时间的推移而变化——这种可变性可能会给现有的有效表示数据的方法带来困难。当前的方法使用较少数量的原始数据的属性或特征来经济地表示数据。标准的特征表示，如傅里叶和小波表示，在表示这些新采集技术的数据时可能效率不高。克服这种不匹配的一个范例是数据驱动方法，其中算法处理数据以学习给定数据的新颖有效的特征表示。虽然这些方法更有用，但这些方法可能无法很好地扩展到大量数据集。这项工作为数据驱动的特征学习设计了新的方法，这些方法对噪声、时变数据具有可扩展性和鲁棒性。它提出了一种“主动筛选”的方法来学习新功能；数据处理算法采用低复杂度标准筛选有用的信息点进行特征学习。主动筛选的优点包括减少计算和存储开销，以及拒绝异常值或其他虚假和误导性数据的能力。研究人员开发了生成模型下数据一致性估计的主动筛选方法，分析了判别字典学习中主动筛选中表示和分类之间的权衡，并将主动筛选分析扩展到分布式字典学习的分布式设置。他们研究了这些方法在170+患者的两个大规模心电图（ECG）数据集上的前景。这项工作结合了统计学（特征筛选）和机器学习（主动学习和选择性抽样）的思想，设计了来自大量数据集的复杂信号的有效表示，并可能通过将筛选思想纳入技术本身，为新的数据采集技术的设计提供信息。

Advances in sensing and data acquisition technologies make it easy to generate vast quantities of data that must be stored, communicated, processed, and understood. This data may be of variable quality and the nature of the data may vary over time -- this variability can cause difficulties for existing approaches to efficiently represent the data. Current methods use economical representations of the data in terms of a smaller number of properties, or features, of the raw data. Standard feature representations such as Fourier and wavelet representations may not be efficient at representing the data from these new acquisition technologies. One paradigm to overcome this mismatch is the data-driven approach, in which an algorithm processes the data to learn a novel and efficient feature representation for the given data. While these are more useful, such approaches may not scale well to massive data sets.This work designs new methods for data-driven feature learning that are scalable and robust to noisy, time-varying data. It proposes an "active screening" approach to learning new features; the data processing algorithm uses a low-complexity criterion to screen for useful and informative points for feature learning. Advantages of active screening include a reduction in the computational and storage overhead as well as the ability to reject outliers or other spurious and misleading data. The investigators develop active screening methods for consistent estimation under generative models for the data, analyze the tradeoff between representation and classification in active screening for discriminative dictionary learning, and extend the active screening analysis to distributed settings for distributed dictionary learning. They investigate the promise of these methods on two large-scale electrocardiography (ECG) datasets of 170+ patients. This work combines ideas from statistics (feature screening) and machine learning (active learning and selective sampling) to design efficient representations of complex signals from massive data sets and may inform the design of new data acquisition technologies by incorporating screening ideas into the technologies themselves.