Dissertation Defense: Venkat Sai Suman Lamba Karanam
Friday, June 6
1:30 PM
347 Avery Hall or Zoom: https://unl.zoom.us/j/92050562209
“Non-Blackbox Robust Design of Machine Learning in Networks”
The saying “let’s not reinvent the wheel” in ML/AI adoption implies that model architecture design be left for pure ML/AI researchers, while network researchers focus on input preprocessing (e.g., formatting the packet data to be fed to a model), hyperparameter fine-tuning and a trial-and-error approach to find the “best” result. This pipeline may work in theory, but is not always practical, especially considering the diversity in network behavior (e.g. communication patterns) in any given network environment (even the same network that was used originally likely behaves differently, eventually). This dissertation is built around the following principle. Understanding the network behavior and designing a mathematical model for that behavior helps reduce the effort and overhead involved in producing satisfactory results with that model. For example, for a traffic prediction problem in a communication network, understanding the traffic patterns and designing a mathematical model that decomposes its key features such as trends, seasonality and cycles can help reduce the ML model’s complexity such as the number of layers, and can help achieve desired accuracy faster. While tackling the mathematical modeling of networks at-large is an impossibly daunting task and cannot be addressed in any single work, we sought out to address the task across multiple arcs. Intelligent techniques with ML/AI can help predict transfer patterns, resource usage, attack detection and even network infrastructure planning. For any semblance of network intelligence, modeling the statistical properties of the data are essential, either at micro- or macro-levels, and in the short- or long-term. This dissertation investigates the importance of careful mathematical modeling before/while adopting ML/AI techniques. We show that modeling the communication network behavior (say, traffic patterns for a traffic forecasting model) using a mathematical model (say, PDE or ODE) is more reliable than adopting a published ML model even if it promises great results in its original publication. Such effort produces objective, optimization and loss functions while keeping the ML/AI model architecture close to baseline. We chose to present proof-of-concepts across several selected problem statements inherent to the networked systems, namely, network management, large-scale network analysis, security and traffic classification, and traffic prediction. Each separate work in this dissertation focuses on one of these problems.
Committee:
Prof. Byrav Ramamurthy, Advisor
Prof. Lisong Xu
Prof. Massimiliano Pierobon
Prof. Yi Qia (External Member)