EAGER: Leveraging Structure to Realize the Promise of Transfer Learning

统计机器学习的成功关键在于能够访问大量数据进行训练。学习算法在数据贫乏的情况下变得不那么有效。这些挑战的例子包括从图像中识别不常见的视觉类别，理解文本和音频语料库都很昂贵的稀有语言，使辅助机器人适应新的环境和主人，并识别罕见的疾病。 迁移学习已经成为一个有吸引力的框架，以解决数据不足的挑战。迁移学习背后的基本思想是利用一组相关任务，其训练数据丰富，以帮助学习目标任务。该研究项目有几个更广泛的影响。迁移学习的最新进展将与PI的教学和研究活动相结合，面向来自不同科学背景的研究生和本科生。该项目将积极吸引本科生参与研究。计划中的研究成果将通过教程、评论文章/调查、特邀演讲和开源软件迅速广泛地传播给科学界。尽管取得了进展，迁移学习方法在很大程度上限于分类任务，其目标是学习表示为欧几里得空间中的点的数据样本的标记函数。 相比之下，许多应用问题中的数据结构复杂、丰富。 例子包括复杂的视觉场景，其中对象类别之间存在很强的上下文依赖性，以及多模态数据，其中每种模态都与其他模态互补。有效地利用这些数据中的依赖性和结构，可能会提高迁移学习相对于忽略它们的方法的有效性。该项目开发了结构化迁移学习的统计方法，并应用于计算机视觉和机器人技术中的问题。 该项目侧重于两个方向：（1）结构化预测问题的迁移学习，以及（2）跨模态迁移学习。 该研究开发了新的统计学习方法，加深了理解，并发明了实用的统计算法来解决复杂类型数据的迁移学习问题。 其次，将本发明的方法应用于计算机视觉和机器人感知中的实际应用问题。该项目将表明，数据结构的适当性推进了智能和自主系统在感知复杂和具有挑战性的现实世界环境方面的最新发展。

The success of statistical machine learning relies critically on having access to a large amount of data for training. Learning algorithms become much less effective in data-poor situations. Examples of such challenges are recognizing uncommon visual categories from their images, understanding rare languages where both text and audio corpora are expensive to collect, adapting and personalizing assistive robots to new environments and owners, and identifying rare forms of diseases. Transfer learning has been emerging as an appealing framework to address the challenge of being poor in data. The essential idea behind transfer learning is to leverage a cohort of related tasks, whose training data are abundant, to help to learn target tasks. The research project has several broader impacts. The most recent advances in transfer learning will be incorporated and integrated with the PI's teaching and research activities for graduate and undergraduate students from diverse scientific backgrounds. The project will actively engage undergraduate students in research. The results of the planned research will be rapidly and broadly disseminated to scientific communities via tutorials, review articles/surveys, invited talks and open-source software.Despite progress, transfer learning methods are largely limited to classification tasks where the goal is to learn a labeling function for data samples represented as points in the Euclidean space. In contrast, data in many application problems are complex and rich in structure. Examples include complex visual scenes where there are strong contextual dependency among object categories, and multimodal data where each modality is complementary to the others. Effectively exploiting the dependency and structures in such data will likely improve the effectiveness of transfer learning relative to methods that ignore them. This project develops statistical methods for structured transfer learning, with applications to problems in computer vision and robotics. The project focuses on two directions: (1) transfer learning for structured prediction problems, and (2) cross-modal transfer learning. The research develops new statistical learning methods that deepen understanding and invents practical statistical algorithm to tackle transfer learning problems for data with complex types. Secondly, the invented methods are applied to practical applications problems in computer vision and robotic perceptions. The project will show that proper of structure in data advances the state-of-the-art of intelligent and autonomous systems in perceiving complex and challenging real-world environments.