III: Small: NeuroDB: A Neural Network Framework for Efficiently Answering Database Queries Approximately

我们正在见证一个数据爆炸的时代。人、传感器和计算机不断产生大量数据，然后存储在计算机系统中。许多现代应用程序要求这些系统快速、准确地回答有关数据的问题，并且存储成本低。这是一项具有挑战性的任务，因为提高速度是以牺牲准确性为代价的，并且在三个目标（速度、存储开销和准确性）之间存在权衡。解决这些问题的传统方法是开发定制的计算机算法，每个算法通过在这些目标之间达成妥协来回答特定类型的问题（称为查询类型）。事实证明，这样的过程非常耗时，需要对每种查询类型重复执行。最终目标是设计一个系统，该系统可以在可能的算法中自动搜索，以便为每种查询类型找到具有最佳时间/空间/精度权衡的算法。这个项目在这个方向上迈出了第一步，利用人工智能（AI）用一个神经网络系统（类似于人类大脑中的神经元网络）取代所有针对不同查询类型的定制算法，这个神经网络系统可以在看到足够的问题和答案示例后回答任何查询类型。这种方法将显著减少（或消除）针对不同查询类型设计/选择算法时的人为干预，同时改善回答这些问题的时间/空间/精度权衡。这将使广泛的应用程序领域受益，特别是考虑到当今大多数应用程序的数据驱动性质。例如，这种方法将极大地有利于数据驱动的领域，如智能城市、交通、能源、环境和健康。最后，这个项目也将对不同的学生（本科生和研究生）非常有吸引力，这些学生现在对人工智能更感兴趣，并将招募他们为数据管理领域做出贡献。这个项目的主要论点是查询可以用近似的函数来表示。因此，将开发一个通用的神经网络框架，称为NeuroDB，它可以学习一个模型来近似查询函数。然后，学习的模型可以用来找到任何类型查询的近似答案。NeuroDB最初的重点将放在距离到最近邻查询和范围聚合查询上，这是许多实际应用程序的两个重要组成部分。初步结果表明，对于这两种查询类型，与最先进的手工算法相比，NeuroDB的速度要快几个数量级，并且只使用一小部分数据大小用于存储，同时提供相似的准确性（模型简洁地学习查询和数据分布）。然后，NeuroDB将扩展为一个通用框架，可以回答更广泛的查询类型集，从而用神经网络取代不同的数据库操作。这样做的挑战是双重的。首先，即使是为回答特定的查询类型而设计一个神经网络也是具有挑战性的。它需要从理论上和经验上对神经网络的近似能力有更深的理解，以及它如何随着网络架构和训练过程而变化，对于特定的查询类型。其次，需要对特定查询类型的观察结果进行概括，以设计一个可以回答其他查询类型的框架，并且需要一种学习方法，能够训练准确的模型来回答这些查询类型。NeuroDB将作为一个开源软件系统开发，供研究界使用和扩展。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。

We are witnessing the age of data explosion. Large amounts of data are continuously generated by people, sensors and computers and then stored in computer systems. Many modern applications require these systems answer questions about the data quickly, accurately and with low storage costs. This is a challenging task, because increasing speed comes at the expense of accuracy, and there is a trade-off between the three goals (speed, storage overhead and accuracy). The traditional approach to address these issues was to develop customized computer algorithms, each designed to answer a specific type of question (called query type), by striking a compromise between these goals. Such a process has proven to be time consuming and needs to be repeated for each query type. The ultimate goal is to design a system that can automatically search among possible algorithms to find the ones with the best time/space/accuracy trade-offs for each query type. This project takes a first step in this direction by utilizing Artificial Intelligence (AI) to replace all customized algorithms for different query types with a single neural network system (similar to network of neurons in human brain) that can answer any query type after seeing enough examples of questions and answers. Such an approach will significantly reduce (or eliminate) human intervention in designing/choosing algorithms for different query types, while improving upon the time/space/accuracy trade-offs for answering them. This will benefit a broad range of application domains, especially given the data-driven nature of most of today's applications. For example, this approach will greatly benefit data-driven areas, such as smart-cities, transportation, energy, environment and health. Finally, this project will also be extremely attractive to diverse students (both undergraduates and graduates) who are these days more attracted to AI, and will recruit them to contribute to the field of data management.The main thesis of this project is that queries can be represented by functions which can be approximated. Thus, a generic neural network framework will be developed, dubbed NeuroDB, that can learn a model to approximate the query function. The learned model can then be used to find approximate answers to any type of query. The initial focus of NeuroDB will be on distance-to-nearest-neighbor queries and range aggregate queries, two important building blocks of many real-world applications. Preliminary results show that for these two query types, NeuroDB is orders of magnitude faster compared to state-of-the-art hand-crafted algorithms, and only uses a fraction of the data size for storage, while providing similar accuracy (the model concisely learns query and data distributions). NeuroDB will then be extended to a generic framework that can answer a broader set of query types, thus replacing different database operations with neural networks. The challenges in doing so are twofold. First, designing a neural network even for answering a specific query type is challenging. It requires a deeper understanding, both theoretically and empirically, of the approximation power of neural networks, and how it changes with network architecture and training process, for the specific query type. Second, the observations about specific query types need to be generalized to design a framework that can answer other query types, and a learning methodology is needed that is able to train accurate models for answering them. NeuroDB will be developed as an open-source software system to be used and extended by the research community.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.