Research
Our research spans AI systems engineering, bias propagation in machine learning, software reliability, and energy-efficient scheduling — united by a commitment to building AI systems that are reliable, fair, and rigorously evaluated.
AI Bias & Fairness
Investigating how bias originates, propagates, and compounds through machine learning pipelines — and building frameworks to detect, audit, and correct it at scale.
AI Systems & IoT
Designing adaptive, energy-efficient scheduling and resource management strategies for AI systems operating under real-time and resource-constrained conditions.
A central focus of the lab is understanding how bias originates, amplifies, and compounds as data moves through successive stages of a machine learning pipeline — from data collection and preprocessing, through model training and evaluation, to deployment.
Our work introduces a state-based framework for modeling stage-aware bias propagation, and investigates the interplay between bias and uncertainty across pipeline stages. We identify distinct regimes in which bias is amplifiable and detectable versus hidden and self-reinforcing — with significant implications for audit methodology and model governance.
Responsible AI requires more than post-hoc fairness auditing — it demands principled evaluation methodologies built into the AI development lifecycle. This research area develops frameworks for assessing the fairness, reliability, and societal impact of AI systems, drawing on both formal methods and empirical analysis.
Our approach is informed by Dr. Nam's experience as a Patent Examiner at the USPTO, where she evaluated AI and software system applications for novelty and non-obviousness — providing a unique lens on how AI capabilities are technically characterized and legally bounded.
AI systems are software systems, and their reliability properties must be rigorously studied. This research examines failure modes, testability, and verification strategies in software systems that underpin modern AI applications — with a particular focus on automated test data generation and data dependency analysis.
Early work in this area produced publications on data dependence-based testability transformation and automated test generation (ISSRE 2005, SEA 2003). This thread is now being reconnected with the bias propagation work, as data dependencies are a key mechanism through which pipeline-stage bias can be traced and audited.
Modern AI deployments face highly variable workload profiles — inference requests surge unpredictably, and real-time applications demand consistent latency guarantees. This research develops adaptive scheduling and resource allocation policies that dynamically respond to workload fluctuations without sacrificing performance or correctness.
Our approaches combine processor-level dynamic voltage scaling (DVS) with intelligent memory allocation strategies, producing schedulability analyses and resource planning policies validated on both simulated and real IoT workloads. This work has resulted in a granted patent and a series of publications in IEEE conference proceedings and journals.
Energy efficiency is a first-class concern in edge AI and IoT deployments. This research develops schedulability analyses and task scheduling methods that simultaneously address real-time deadline constraints and energy consumption by coordinating processor voltage/frequency scaling with memory allocation strategies.
Our work on non-volatile memory (NVRAM) integration with real-time scheduling produced a tight schedulability analysis published in Micromachines (2019), and earlier work explored memory allocation and processor voltage scaling for energy-efficient IoT scheduling (IEEE/ACIS ICIS 2017).