BIGDATA: F: Compositional Learning, Maps and Transfer: Statistical and Machine Learning on Collections of Data Sets

人类智力的标志之一是不仅能够找到难题的解决方案，而且能够从过去的经验中学习和积累知识，这些知识可能（部分地）转移到快速解决新问题。该项目将从数据集和机器学习问题的集合中开发新的学习组合规则的基础技术。研究者将开发的构建模块能够跨多个数据集和模式共享学习。第一个构建块将使机器学习算法能够存储过去问题的解决方案，并使用地图和抽象将知识转移到新问题上。这需要学习地图的有效技术，如何组合它们以实现知识转移，所有这些都以与问题及其解决方案的表示相兼容的方式进行，这也需要自动学习。这些想法将在一系列问题上进行测试，从图像的对象和模式识别到交互代理系统的行为，从融合不同传感器获得的数据集到控制虚拟和真实代理。该项目将提供机器学习的一般基础结果，可以应用于几乎任何人类努力领域的应用。研究者将开发专注于表征和迁移学习的新技术，特别是：(i)作文学习；通过数据集之间的组合映射以及数据集上的函数（用于分类和回归任务）来学习和分解的能力（例如，任务f可以通过使用到一个已经发生学习的数据集的映射h和该数据上已经学习的函数g来学习），以提高学习率，知识提取和跨数据集和数据类型的转移；地图学习：有效地学习、表示、储存、回忆和应用可能具有不同模式的复杂数据集之间的地图的能力；但也要学习将一个任务转化为另一个任务，并将知识从一个任务转移到另一个任务的地图；（iii）表示学习：学习如何有效地表示、存储和召回复杂数据集的能力，跨越多个传感器模式，跨越不同的抽象层次——例如，学习来自多种类型传感器的数据的有效表示，学习分类器和回归函数，或学习基于代理的系统中的交互核，以及在传感器模式、数据集、动态系统之间传递这些功能。在推进这些学习能力的最新技术的同时，该研究将解决在学习不变性和执行图像中的对象识别任务方面的应用，检测图像中的对象是新的还是已知的，通过观察交互代理系统的轨迹来学习交互规则，并在学习系统的背景下实现组合学习的思想，包括虚拟的(例如，使用OpenAI挑战)和真实的（例如，使用机器人），在难度不断增加的任务序列上。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

One of the landmarks of human intelligence is the ability to not only find solutions to hard problems, but to learn from past experiences and accumulate knowledge that may be (partially) transferred for quickly solving new problems. This project will develop novel foundational techniques for learning compositional rules, from collections of data sets and machine learning problems. The building blocks that the investigator will develop enable sharing of learning across multiple data sets and modalities. A first building block will enable machine learning algorithms to store solutions to past problems and use maps and abstractions to transfer knowledge to new problems. This requires efficient techniques for learning maps, how to compose them to enable knowledge transfer, all in a way that is compatible with the representation of the problems and their solutions, which also need to be automatically learned. These ideas will be tested on problems ranging from object and pattern recognition of images to behavior of interacting agent systems, from fusing data sets acquired with different sensors to controlling virtual and real agents. This project will provide general, foundational results in machine learning, which can be applied to applications in virtually any domain of human endeavor. The investigator will develop new techniques focused on representation and transfer learning, in particular: (i) Compositional Learning: the ability to learn and factorize through composition maps between data sets, and of functions (for classification and regression tasks) on data sets (e.g. the task f may be learned by using the map h to one data set on which learning already occurred and the already-learned function g on that data), in order to enhance both learning rates, knowledge extraction and transfer across data sets and data types; (ii) Map Learning: the ability to efficiently learn, represent, store, recall and apply maps between complex data sets, possibly of different modalities; but also learn maps that transform, at least approximately, one task into another, and transfer knowledge from one task to another; (iii) Representation Learning: the ability to learn how to efficiently represent, store and recall complex data sets, across multiple sensor modalities, and across different levels of abstractions -- for example, learning efficient representations of data from multiple types of sensors, learning of classifiers and regression functions, or learning interaction kernels in agent-based systems, as well as transfer those functions across sensor modalities, data sets, dynamical systems. While advancing current state of art techniques in each of these learning abilities, the research will tackle applications in learning invariances and performing object recognition tasks in images, detecting whether objects in an image are new or known, learn interaction rules from observing trajectories of interacting agent systems, and implement the ideas of compositional learning in the context of learning systems both virtual (for examples, using the OpenAI challenges) and real (for example, using robots), on sequences of tasks of increasing difficulty.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.