CDS&E-MSS: Sparsely Activated Bayesian Neural Networks from Deep Gaussian Processes

这个计算和数据支持的数学和统计科学科学与工程（CDS&E-MSS）项目旨在开发一种系统的方法来构建深度贝叶斯神经网络，该网络既具有计算效率，又适合模型设计。该项目的结果预计将在准确性、信心和效率方面显著提高，超过当前的机器学习和人工智能模型。开源软件将在这个项目的过程中开发，为研究人员和从业者提供可访问的工具。预计该研究将对自动驾驶汽车、医疗诊断、交易金融等广泛应用产生影响。研究人员还致力于促进少数民族和代表性不足群体的STEM教育和研究机会，该项目还为研究生提供研究培训机会。该研究预计将推动深度高斯过程模型和相关贝叶斯神经网络在几个关键领域的发展，包括理论和算法的发展，它们在生成任务中的应用，以及软件的开发和验证。研发计划主要由以下三个部分组成：1)更好地理解深度张量马尔可夫高斯过程模型的性质，构建更广泛的模型，这些模型具有更复杂的结构，但仍然可以被稀疏结构的深度贝叶斯神经网络近似；2)探索深度高斯过程模型的随机性，创建新的条件生成模型；3)开发软件工具，并使用来自最先进的无线通信和计算机视觉系统的真实数据集验证项目结果。该奖项由美国国家科学基金会数学科学部颁发，由土木、机械和制造创新部联合支持。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。

This Computational and Data-Enabled Science and Engineering in Mathematical and Statistical Sciences (CDS&E-MSS) project aims to develop a systematic approach for constructing deep Bayesian neural networks that are both computationally efficient and amenable to model designs. The results of this project are expected to lead to significant improvements in terms of accuracy, confidence, and efficiency over the current machine learning and artificial intelligence models. Open-source software will be developed during the course of this project, providing accessible tools for researchers and practitioners. This study is expected to impact a wide range of applications including self-driving cars, medical diagnostics, and trading and finance. The investigators are also committed to promoting STEM education and research opportunities for minority and underrepresented groups, and this project also provides research training opportunities for graduate students. The research is expected to advance deep Gaussian process models and relevant Bayesian neural networks in several key areas, including theory and algorithmic development, their applications to generative tasks, and software development and validation. The research and development plan mainly consists of the following three components: 1) to better understand the properties of deep tensor Markov Gaussian process models and to construct a broader class of models that have more complex architectures but can still be approximated by deep Bayesian neural networks with sparse structure; 2) to explore the stochastic nature of deep Gaussian process models to create novel conditional generative models; and 3) to develop software tools and to validate the project outcomes using real-world datasets from the state-of-the-art wireless communication and computer vision systems.This award by the NSF Division of Mathematical Sciences is jointly supported by the Division of Civil, Mechanical, and Manufacturing Innovation.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.