Autonomous decision making in scientific exploration

观测科学有一个问题。我们知道的太多了。电子技术的最新进展，特别是在相机和光谱传感器设计方面，使收集前所未有的大量信息成为可能。不幸的是，收集和存储传感器信息的能力远远超过了分析这些信息并从中得出有意义的科学推论的能力。在大多数地面环境中，这个问题只是令人恼火。希望对世界偏远地区进行观测的科学家要么必须亲自前往那些偏远地区（这既昂贵又危险），要么必须煞费苦心地仔细研究由遥感器获得的大量数据，从科学的角度来看，这些数据中的大多数并不有趣，因为它们要么没有对感兴趣的现象进行观测，要么只是概括了先前的观测结果。在行星际探索的情况下，还需要额外的费用，称为“数据预算”。即使行星探测器任务变得越来越有能力，有了增强的发电厂和显著升级的传感器套件，行星际距离带来的低带宽通信意味着科学和工程团队被迫进入一种日常的分类谈判，以确定“该轮到谁使用天线”？在数据预算有限的情况下，减少浪费在传输数据上的时间变得更加重要，因为这些数据不包含新信息，所以没有什么价值。简而言之，这是一个经典的工程问题。我们有两种宝贵而有限的资源：训练有素的科学家的认知能力和注意力，以及星际机器人的带宽。我们提出的研究计划的长期目标是开发一套工具，利用计算机视觉和机器学习方面的现代技术，使某些科学任务自动化，以最大限度地提高研究的科学回报。这项研究并没有试图“取代”人类科学家。机器学习领域离制定、推理、测试和讨论假设重要性的能力还很远。然而，我们相信科学界需要先进的工具来增强研究者搜索大型数据集的能力，特别是大型视觉和光谱数据集，以识别感兴趣的特征。我们设想这些工具充当“兴趣过滤器”，筛选tb级的数据，并在观察到某些想要的现象时通知人类。拟议的研究项目是跨学科的。主要的影响将在观测科学（行星科学、地质学、天文学等）中感受到，在这些领域将应用这项工作，加快资料收集和分析的进程，而要处理的大多数具体项目都是由与科学同事合作确定的问题所推动的。

The observational sciences have a problem.  We know too much.  Recent advances in electronics, and especially in camera and spectral sensor design have enabled the collection of unprecedented quantities of information. Unfortunately, the ability to collect and store sensor information have far outstripped the ability to analyze it and draw meaningful scientific inferences from it. In most terrestrial contexts, the problem is merely irritating. Scientists who wish to make observations of remote parts of the world must either physically travel to those remote locations (which is costly and dangerous) or painstakingly pore over large quantities of data obtained by remote sensors, most of which is not interesting from a scientific perspective, because it either contains no observations of the phenomenon of interest or merely recapitulates prior observations. In the context of interplanetary exploration, there is an additional cost imposed, which is called the "data budget". Even as planetary rover missions become more and more capable, with enhanced power plants and significantly upgraded sensor suites, the low bandwidth communications imposed by interplanetary distances mean that science and engineering teams are forced into a sort of daily triage negotiation to determine "whose turn is it to use the antenna"? It becomes even more important, in the context of a limited data budget, to reduce the wasted time spent transmitting data that, since it contains no new information, has very little value. In short, here is a classic engineering problem. We have two precious, limited resources: the cognitive capability and attention span of trained scientists, and interplanetary robot bandwidth. The long term goal of our proposed research program is is to develop a suite of tools, taking advantage of modern techniques in computer vision and machine learning, that can automate certain scientific tasks to maximize the scientific return from investigation. There is no attempt in this research to "replace" human scientists. The field of machine learning is very far from the ability to formulate, reason about, test and discuss the importance of hypotheses. However, we believe there is a need in the scientific community for advanced tools that enhance an investigator's ability to search a large dataset, especially large visual and spectral datasets, to identify features of interest. We envision these tools acting as "interest filters", sifting through terabytes of data and notifying a human when some desired phenomenon is observed. The proposed research program is interdisciplinary. The primary impacts will be felt in the observational sciences (planetary science, geology, astronomy, etc) where the work will be applied, by accelerating the process of information collection and analysis, and most of the specific projects to be addressed are motivated by problems identified in collaboration with scientific colleagues.