Distributed Machine Learning for Automatic Annotation and Analyses of Vast Distributed Image Archives

该博士项目旨在解决如何使用分布式机器学习技术(如联合学习)有效分析海量高分辨率图像的内容这一挑战。这些图像是由新兴的传感器生成的，并存储在地球上的各个位置。如果可以将所有图像集中到一个中心位置，则可以使用集中式机器学习对图像进行自动注释，从而生成多种类型(汽车、树木、建筑物、桥梁等)中的每一种对象的地理时间局部化列表。然后可以以这样的方式处理该列表，例如，可以标识改变和/或对查询对象进行本地化的类似对象。由于图像单独很大，而且通常数量也很大，通信带宽考虑意味着将所有图像发送到一个集中位置是不切实际的。关于本地数据集的隐私可能使得传输到中央位置在法律上是不可行的。所需要的是一种机制，通过该机制，可以在(真实的)分布式、异步设置中模拟(假设的)集中式流程。这将需要开发促进分布式存储和查询的基础设施(例如，通过分布式空间索引、支持学习的智能查询等)以及开发可以在这种分布式、异步设置中操作的分布式算法(例如，通过地图缩减、联合学习等)。该项目旨在创建提供分布式机器学习机制的重要数据科学和计算解决方案，从而可以在(真实)分布式设置中模拟(假设的)集中过程。这将需要开发促进分布式存储和查询的基础设施(例如，通过分布式空间索引、支持学习的智能查询等)以及开发能够在这种分布式、异步设置中操作的分布式算法(例如，通过地图简化、联合学习等)。

This PhD project is to tackle the challenge of efficiently analysing the content of vast volumes of high-resolution imagery with distributed machine learning techniques such as federated learning. The images are generated with emerging sensors and stored in locations that span the Earth. Were it possible to bring all the imagery to one central location, it would be possible to use centralised machine learning to auto-annotate the imagery and thereby generate a list of geo-temporally localised objects of each of many types (cars, trees, buildings, bridges etc). This list could then be processed in such a way that, for example, changes could be identified and/or similar objects to a query object localised. Since the images are individually large and typically also very large in number, communications bandwidth considerations mean that it is not practical to send all the imagery to one centralised location. The privacy regarding local datasets may render the transmission to a central location legally infeasible. What is needed is a mechanism whereby the (hypothetical) centralised processes can be emulated in a (real) distributed, asynchronous setting. This will require development of both the infrastructure to facilitate distributed storage and querying (via e.g., distributed spatial indexing, learning-enabled intelligent querying, etc) as well as development of distributed algorithms (via e.g., map-reduce, federated learning, etc) that can operate in such a distributed, asynchronous setting.The project aims to create non-trivial data science and computing solutions that offer a distributed machine learning mechanism whereby the (hypothetical) centralised processes can be emulated in a (real) distributed setting. This will require development of both the infrastructure to facilitate distributed storage and querying (via e.g., distributed spatial indexing, learning-enabled intelligent querying, etc) as well as development of distributed algorithms (via e.g., map-reduce, federated learning, etc) that can operate in such a distributed, asynchronous setting.