Reasoning Under Uncertainty
Bayesian deep learning methods that know when to say "I don't know" — applied to medical triage and autonomous navigation.
View related papers ↓DEPARTMENT OF COMPUTER SCIENCE — INSTITUTE OF TECHNOLOGY
Professor of Artificial Intelligence & Distributed Systems
Building machines that reason under uncertainty, and the students who'll out‑build me. Eighteen years in front of a chalkboard, fourteen in front of a GPU cluster.
Dr. Voss leads the Reasoning Systems Lab, where her team studies how learning systems can make trustworthy decisions when the data runs out. Before joining the faculty, she spent four years building inference infrastructure in industry — work she now mines for examples in every lecture.
She believes the best research questions are the ones a curious undergraduate asks by accident, and she keeps a running notebook of them on her office door.
Bayesian deep learning methods that know when to say "I don't know" — applied to medical triage and autonomous navigation.
View related papers ↓Training and serving models across thousands of edge devices without ever centralizing raw data.
View related papers ↓How people build (and lose) trust in AI recommendations, and how interfaces can communicate model confidence honestly.
View related papers ↓Making model training accessible on modest hardware for labs and classrooms without datacenter budgets.
View related papers ↓We introduce a method for learned models to defer to clinicians when predictive confidence falls below a task-specific threshold, reducing false-confidence errors by 41% on three hospital triage datasets without sacrificing throughput.
A longitudinal deployment across 4,000 edge devices shows that periodic re-synchronization, rather than continuous federated updates, yields better accuracy-per-watt for resource-constrained classrooms and clinics.
A user study of 312 participants finds that explanation length matters less than whether the system discloses its own uncertainty — a result we use to redesign confidence indicators in three production systems.
A curriculum of gradient checkpointing, sparse attention, and staged distillation that brings transformer pretraining within reach of a single consumer GPU, with full lab materials released for instructors.
A drift-aware particle filter that maintains calibrated uncertainty estimates over multi-hour autonomous navigation tasks, tested on six campus delivery robots.
Graduate seminar on Bayesian inference, variational methods, and uncertainty quantification in deep networks.
FALL · SPRINGCore undergraduate course covering consensus, replication, and the federated systems used throughout the lab's research.
FALLFirst-year gateway course — the one Dr. Voss has refused to stop teaching for eighteen years running.
FALL · SPRINGWeekly seminar for PhD candidates presenting work-in-progress to the Reasoning Systems Lab and invited critics.
YEAR-ROUNDAwarded by the Institute's Faculty Senate for sustained excellence in undergraduate instruction.
Recognized for sustained contributions to distributed machine learning systems.
Five-year grant supporting research into calibrated uncertainty for safety-critical AI systems.
Redesigned the department's graduate admissions process to weight research potential over standardized testing.
For early work on drift-aware uncertainty estimation in mobile robotics.
Office hours are real and held weekly — students, collaborators, and curious sophomores are equally welcome.