[CS] Zhujun Xiao Candidacy Exam / March 16, 2021

[pbaclawski at uchicago.edu]

This is an announcement of Zhujun Xiao's Candidacy Exam.
===============================================
Date:  March 16, 2021

Time: 4:00PM CST

Location: Via zoom

https://uchicago.zoom.us/j/91566872732?pwd=T2VmUzEvdVpNekxhZGpmaVNNbjJTdz09
Meeting ID: 915 6687 2732
Passcode: 711174

Candidacy Candidate: Zhujun Xiao

Title: Towards Large-Scale Deployment of Machine Learning Systems

Abstract: While ML researchers often design and evaluate ML models based on existing datasets, real-world deployment of these ML models must address system-level challenges. My research considers the problem of deploying ML models in large-scale systems and addresses two system-level challenges: heterogeneity and scalability. In this thesis proposal I will present my recent works on designing and evaluating ML models for large-scale deployments. 

I will start from presenting a distributed spectrum monitoring framework for wide-area cellular networks, where deep neural network (DNN) models are deployed to detect spectrum usage anomalies on the fly. These DNNs are distributed to spectrum observers that are either static or mobile. The key challenge facing this system is the complex, heterogenous, and time-varying spectrum observations seen by each observer, caused by inherent radio propagation dynamics and diverse mobility patterns.  Such complexity makes it hard to train and deploy accurate spectrum anomaly detection models at scale.  My work addresses this challenge by leveraging the hierarchical network structure of cellular networks. We build a single context- and mobility-agnostic DNN model for any observer within a single cell and ``customize’’ this model across cells via transfer learning.

Next, I consider the problem of evaluating large-scale ML systems to reveal their performance on real-world workloads. Here I focus on ML-based video analytic pipelines (VAPs), given their increasing popularity in smart city monitoring. While many VAPs have been developed to address the tradeoff between accuracy and resource cost, our measurement shows that existing VAP evaluations are incomplete, often producing premature or ambiguous results. This is because existing evaluations only test VAPs on video datasets sampled from target scenarios, which fail to capture the impact of content dynamics commonly seen in real-world deployments. This motivates us to build Yoda, the first VAP benchmark to comprehensively evaluate  VAPs using carefully curated video dataset. Yoda evaluates each VAP by characterizing its accuracy/cost tradeoff while also revealing its dependence on video content characteristics.

Finally, I will briefly discuss ongoing/future work, and my thesis plan.

Advisors: Ben Zhao and Heather Zheng

Committee Members: Junchen Jiang, Heather Zheng, and Ben Zhao