BIGDATA: F: DKA: CSD: Human and Machine Co-Processing

人类专家对数据分析至关重要。它们的作用包括筛选大型数据集，以促进搜索，检索和机器学习。在这些任务中，人类通常比机器表现得更好，但人类专家的速度和能力是人机协同处理的限制因素。 该项目解决了人机协同处理的两个方面：筛选大数据以产生可管理的子集用于人类专家分析，以及机器学习算法，以最少的人类交互有效地从人类专家那里学习。 这有着广泛的应用范围;为了确保结果的广泛适用性，该项目正在评估多个领域的技术：认知科学，大规模天文数据分析和材料科学的实验设计。用于数据筛选的方法基于开发预测模型和识别不适合模型的数据。 一个关键的研究挑战是非平稳环境：基础模型随着时间的推移而变化。 初步工作表明，从有限的一组模型选择的承诺，和新的工作研究更灵活的参数模型。主动学习任务使用多臂强盗问题来建模识别哪些特征对人类决策影响最大。 这项任务还研究了从比较/排名而不是预测中学习;证明可能存在管理人类推理和决策的低维结构，使得学习比其他情况下需要的比较少得多。 这两项任务的共同主题是确保计算复杂度足够低，以便在数据量很大的情况下与人类专家进行实时交互。这是实现使用有界近似和凸松弛的优化程序用于指导的相互作用。

Human experts are crucial to data analysis. Their roles include sifting through large datasets to facilitate search, retrieval, and machine learning. Humans often perform much better than machines at such tasks, but the speed and capacity of human experts is a limiting factor in the human-machine co-processing. This project is addressing two aspects of human-machine co-processing: winnowing Big Data to produce manageable subsets for human expert analysis, and machine learning algorithms that learn efficiently from human experts with a minimal amount of human interaction. This has a wide range of applications; to ensure broad applicability of the results the project is evaluating the techniques in multiple domains: cognitive science, large-scale astronomical data analysis, and experimental design in materials science.The approach used for data winnowing is based on developing predictive models and identifying data that does not fit the models. A key research challenge is non-stationary environments: the underlying model changes over time. Preliminary work shows promise on selection from a finite set of models, and new work investigates more flexible parametric models. The active learning task uses the multi-armed bandit problem to model identify which features have the greatest impact on human decisions. This task also investigates learning from comparisons/rankings rather than predictions; conjecturing that there may exist low-dimensional structure governing human reasoning and decision-making that enables learning with significantly fewer comparisons than might otherwise be required. A common theme in both tasks is ensuring computational complexity is low enough to facilitate real-time interactions with human experts in spite of the volume of data. This is achieved using bounded approximations and convex relaxations of the optimization programs used to guide the interaction.