CAREER: Co-evolution of Machine Intelligence and Continuous Information

点击翻译按钮获取中文摘要

The real world is complex and constantly changing. Information becomes progressively available over time via continuous stream of noisy data. For example, molecular biology databases are always updated with various new gene-protein relationships (e.g., gene regulations, genetic interactions) validated in wet lab experiments. High-energy experiments in astro-particle physics give rise to large amounts of data in the form of continuous high-volume streams. Interpersonal interaction behaviors (e.g., facial expressions, gestures) contain rhythms that are not only correlated in time but also exhibit phase synchronization in an ongoing flow of mutual influence. In order to fully exploit these non-stationary streams of data online and in real-time, this project aims to build machine intelligence capable of quantifying uncertainty, constantly accommodating new information, and consolidating previously acquired knowledge in the meantime. The project is specifically motivated by applications to computational biology, astro-physics, and human interaction behaviors, and will result in innovative online inference methods that are broadly applicable across data-intensive domains. Furthermore, this research will support the interdisciplinary development of a diverse student body of disability, women, and minorities, and the development of graduate-level machine learning courses. The outreach work will contribute to increasing partnerships between academia and industry, and increasing public scientific literacy.The overarching research objective is to create a unified dynamic statistical inference framework based on Bayesian nonparametrics to jointly update deep neural network parameters, adapt neural architectures, and consolidate acquired knowledge. The investigator highlights three inter-related projects that significantly advance this research agenda: (1) neural parameter online inference: design and develop online inference algorithms to recursively update posterior distributions of neural network parameters with continuous data stream, (2) neural architecture online inference: create modeling and computational methods to enable deep neural architectures to automatically go through a qualitative growth to accommodate progressively available information from new data as they are streaming, and (3) dynamic knowledge distillation: design and develop modeling methods to merge old knowledge with new one while both neural parameters and architectures are evolving to extract statistical structures from heterogeneous data stream. These research aims will be complemented by multidisciplinary collaborations to rigorously validate and generalize the framework. This research will lead to efficient statistical online inference methods for the whole information integration and data science.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.