ABI Innovation: Interactive Learning Tools For Individual Identification in Large Biological Image Databases

本研究的目标是在大型生物数据库中开发可扩展的个体识别方法，并将其应用于世界各地正在进行的保护工作中的多个物种。这项工作将众包与计算学习结合起来。在从用户输入中学习的识别系统中，多个输入可以提高识别的质量和速度，从而大大减少任何人工执行的后续工作周期。假设是计算学习将提高人机耦合系统解决方案的质量，以帮助解决大型和多样化的识别问题。在生物图像数据库索引和检索的各个阶段，提出了几种交互式学习算法。该算法基于非参数贝叶斯推理，并提供增量在线学习方法。为了测试互动学习假设，麻省理工学院的单回路系统将被扩展到包括提议的互动学习工具。Sloop是一个强大的视觉工具工具包，已经在多个物种上进行了测试，并支持操作部署。作为本研究的一部分，将开发一个索引多个物种的分布式单桅帆船系统。如果不能量化动物迁徙和运动的空间尺度和其他统计数据，就很难准确估计许多珍稀和濒危物种保护工作的有效性。标记是一种已建立的方法，但其有效性有限，因为它通常是侵入性的，不能大量进行。这项研究着眼于是否可以使用存储在数据库中的照片来单独识别多个物种的大量动物。这里提出的“动物生物识别”将研究公民科学家如何将相关判断和其他输入进行众包，判断如何提高计算机的识别性能，以及性能的提高如何减少公民或专家科学家的工作量。由机器学习介导的共生人机交互机制，以及模式索引和搜索的方法也对其他领域产生影响。例如，这里开发的一些工具已应用于地球科学和天气预报。

The goal of this research is to develop scalable methods for individual identification in large biological databases with application to multiple species in ongoing conservation efforts around the world. This work couples crowd-sourcing with computational learning. In an identification system that learns from user inputs, multiple inputs can accelerate the quality and rate of identification, which in turn dramatically reduces subsequent work-cycles any human performs. The hypothesis is that Computational learning will improve the quality of human-machine coupled system's solutions to help solve large and diverse identification problems. Several interactive learning algorithms are proposed in all stages of indexing and search in Biological Image Databases. The algorithms are based on non-parametric Bayesian inference and deliver incremental online learning methods. To test the interactive learning hypothesis, the MIT Sloop system will be extended to include the proposed interactive learning tools. Sloop is a robust toolkit for vision tools that has been tested on multiple species and supports an operational deployment. As part of this research a distributed Sloop system indexing multiple species will be developed. It is difficult to accurately estimate the effectiveness of conservation efforts for many rare and endangered species without an ability to quantify the spatial scales and other statistics of animal migration and movement. Tagging, an established method, is of limited effectiveness because it is often invasive and cannot be conducted in large numbers. This research looks at whether a large number of animals in multiple species can be identified individually using photographs stored in a database. The "Animal biometrics" proposed here will examine how citizen scientists can crowd-source relevance judgments and other inputs, how judgments can improve the computer's identification performance, and how improved performance reduces the citizen or expert scientist's workload. The mechanics of the symbiotic human-computer interaction mediated by machine learning, and the methods by which patterns are indexed and searched also have impacts in other fields. For example, some of the tools developed here have been applied in Geosciences and Weather Prediction.