[CS] Pouya Mahdi Gholami Candidacy Exam/Jun 3, 2025

[via cs]

This is an announcement of Pouya Mahdi Gholami's Candidacy Exam.
===============================================
Candidate: Pouya Mahdi Gholami

Date: Tuesday, June 03, 2025

Time:  4 pm CST

Location: JCL 298

Title: Adaptive Measures for Improving System Performance in the Unreliable Intermittent Edge

Abstract: Capacitor-based, intermittent sensing systems are a recent advancement in embedded edge sensing. Batteryless operation removes the many disadvantages of batteries (e.g. battery replacement, system size, and environmental concerns), making intermittent systems suitable for many domains from small satellites to infrastructure monitoring. However, accurate and timely sensing is challenging given the unreliability and scarcity of ambient energy sources such as solar and radio frequency. This thesis aims to reduce the energy burden of intermittent sensing via adaptive methods despite the challenging energy conditions.

We begin by developing a framework for deploying Adaptive Sampling Algorithms (ASAs) on the intermittent edge. ASAs aim to maximize sensor data accuracy under scarce energy constraints. However, ASAs cannot be directly deployed on intermittent sensors due to conflicting energy availability assumptions. We resolve this issue by buffering signal data in the time domain and monitoring queue dynamics to maximize sensor accuracy and reduce deadline misses. 

Next, we address the limitations of accurate and timely Dense Neural Network (DNN) inference on the intermittent edge. Modern machine learning methods are computationally expensive and only provide results once the computation is complete. Using a top-down methodology, we first improve the efficiency of such models in intermittent devices by an order of magnitude. Next, we introduce an adaptive inference framework that provides a rich accuracy/energy trade-off and internal adaptation to remain robust despite unreliable energy conditions. Using this framework, we increase inference accuracy and reduce deadline violations on intermittent sensors.

Finally, we propose two new hardware-based frameworks to address capacitor limitations. In the first, we develop a unified, dynamic capacitor architecture to efficiently support the various ranges of available energy on the edge. In the second framework, we note that current intermittent runtimes use a simplified energy-harvesting model that faces challenges and incurs many power failures under unreliable energy conditions. We propose a new framework to address these challenges and provide reliable scheduling under intermittent execution.

Advisors: Hank Hoffmann

Committee Members: Pedro Lopes, Sanjay Krishnan, and Hank Hoffmann