Integrating Rate-Distortion Theory and determinantal point processes
This event is in the past.
11:30 a.m. to 12:20 p.m.
Integrating rate-distortion theory and determinantal point processes to diversify learning data samples for traffic monitoring applications
Dr. Abolfazl Razi, Clemson University
It is known that data sample diversity can influence ML algorithm quality. One potential application for this concept is traffic safety analysis. In this talk, I will review our recent work on developing network-level safety analysis that aims to develop a temporal and spatial crash risk map of highways by processing massive imagery collected from the roadside units (RSUs) and correlating it with the crash reports, in the state of Arizona. However, a key bottleneck here and in many other applications is the limited transmission and processing bandwidth of traffic control infrastructure. This motivated us to seek methods that can reduce data accumulation load by the selective transmission of the most informative data samples (images, and video frames). This can be viewed as selecting the most diverse data samples.
A formal approach to measure sample diversity is Determinantal Point Process (DPP). However, it is appropriate for one-shot inference and not gradual sampling for online learning tasks. In this talk, we first review a fundamental relationship between DPP, and the Rate-distortion (RD) theory used to assess the lossy compression of data samples. We will discuss how this connection can be used to design an RD-based valued function to evaluate the diversity gain of data samples. We also will discuss our recent observation that the diversity of data sampled by DPP seems to have a universal trend of phase transition that quickly approaches its maximum point, then slowly converges to its final limits. This means that using DPP presents its highest value only at the beginning of the data accumulation phase. Based on this fact, we have recently proposed a two-stage approach called RD-DPP that uses DPP for robust initialization and prediction inconsistency for sequential sampling. RD-DPP consistently outperforms DPP-based, uncertainty-based, and random selection methods under all sampling budgets. The results will be discussed in comparison to similar methods.
Dr. Abolfazl Razi is an Associate Professor of Computer Science in the School of Computing (SoC) at Clemson University. Prior to joining CU, He received his B.S., M.S., and Ph.D. degrees, all in Electrical Engineering, respectively from Sharif University (1994–1998), Tehran Polytechnic (1999–2001), and University of Maine (2009–2013). Prior to joining Clemson, was an Assistant Professor of Electrical Engineering at Northern Arizona University (2015-2021). He also held two Postdoctoral Positions at Duke University (2013–2014) and Case Western Reserve University (2014–2015), a visiting Scholar position at the University of Maryland (2013), and a Visiting Professor position at the US Airforce Research Laboratory (Summers 2016 and 17). In addition to his academic endeavors, he served in the Wireless Networking and Smart Card industry for 7 years, holding project manager and R&D positions. Some of Razi’s recent research projects include developing learning-based algorithms for autonomous systems, driving safety analysis, forest fire monitoring and modeling using drones, and nano-resolution 3D image reconstruction using digital holography for supply chain security. These projects are financially sponsored by several organizations, including NSF (NeTS, CPS, PFI, CRI, and CRII programs), USDA (AFRI Program), US AFRL (VFRP and AFROS Programs), Arizona Commerce Authority, BMW Research, MIT Lincoln Laboratory, and Los Alamos National Laboratory (accounting for more than $5 million for individual and collaborative projects). The results of his research have been published in about 95 peer-reviewed journal articles and conference papers. He also filed 6 US Patent Applications.
Online: Zoom link
Join via: ID: 921 6321 2490 (Passcode: 626162)