SHF: Small: Knowledge Integrated Data-Efficient Deep Learning

在过去的十年中，深度学习在各种应用领域取得了巨大的成功，这主要是由于大规模计算能力的同时增加以及用于训练具有大量参数的网络的超大数据集的可用性。众所周知，传统的深度学习受到数据稀缺的困扰，也就是说，由于收集数据的成本太高，或者与必要的数量相比，数据集太小，因此用于学习的数据样本太少（甚至没有）。这种情况在真实的生活中很常见，例如罕见疾病诊断，这大大限制了深度学习的适用性。因此，迫切需要在数据非常有限的情况下进行数据高效的深度学习。该项目旨在为许多重要的现实问题提供创新解决方案，如自动驾驶，制造自动化，机器人，监控，医疗保健和大数据分析。它将支持“人工智能为美国人民”的国家倡议，加速人工智能的发现，并保持美国在人工智能技术方面的领导地位。该项目专注于知识集成的数据高效深度学习，沿着三个不同的方向：（1）大规模知识图谱的数据高效深度学习。知识图作为可靠推理的先验知识被广泛使用，可以用来显著减少训练数据的需求。从结构化学习和最优传输的角度提出了一种新的双层优化方法来解决这个问题。优化中的数学基础也可以更好地理解一些理论问题，例如数据有效学习中的模型复杂性。(2)自我监督的数据增强。数据本身是主观的，可以解释为经验知识。为了解决这个问题，提出了使用自监督的非线性拓扑降维来生成用于训练深度学习模型的新数据。(3)用于数据高效深度学习算法的硬件架构。数据高效的深度学习有可能彻底改变人工智能在建模、预测和决策方面的应用。沿着软件，一个基于FPGA的高能效，可重构的硬件架构被提出用于开发的算法，并将在自动驾驶的感知系统中进行测试。该奖项反映了NSF的法定使命，并通过使用基金会的智力价值和更广泛的影响审查标准进行评估，被认为值得支持。

Over the last decade, deep learning has demonstrated great success in a variety of application areas mainly due to the simultaneous increase of massive computing power and the availability of extremely large datasets for training networks with a large number of parameters. It is also well-known that conventional deep learning suffers from data scarcity, namely, too few (even no) data samples exist for learning because it is too expensive to collect data or the data set is small as compared to the necessary amount. Such cases are common in real life such as rare disease diagnosis which significantly restricts the applicability of deep learning. Thus, there is a crucial need for data-efficient deep learning with very limited data. This project seeks to provide innovative solutions to many important real-world problems such as autonomous driving, manufacturing automation, robotics, surveillance, healthcare, and big data analysis. It will support the national initiative of “Artificial Intelligence for the American People” that accelerates AI discoveries and maintains American leadership in AI technologies.This project focuses on knowledge-integrated data-efficient deep learning, along three different thrusts: (1) Data-efficient deep learning with large-scale knowledge graphs. Knowledge graphs, widely-used as a priori knowledge for reliable reasoning, can be used to significantly reduce the need of training data. A novel bilevel optimization approach is proposed to address this problem from the perspectives of structured learning and optimal transport. The mathematical foundation in optimization can also lead to better understanding of some theoretical questions such as the model complexity in data-efficient learning. (2) Self-supervised data augmentation. Data itself is subjective and open to interpretation as empirical knowledge. To address this problem, nonlinear topological dimension reduction using self-supervision is proposed to generate new data for training deep learning models. (3) Hardware architecture for data-efficient deep learning algorithms. Data efficient deep learning has the potential to revolutionize the applications of AI regarding modeling, predicting, and decision making. Along with software, an FPGA-based power-efficient, reconfigurable hardware architecture is proposed for the developed algorithms and will be tested in a perception system for autonomous driving.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.