Jesse Roberts

Jesse Roberts

Native Tennesseean Researching AI, Language, and Robotics

Research

I work on understanding how neural networks, and language models in particular, learn and reason, and on putting that understanding to work where it matters: keeping endangered languages alive, and detecting tornadoes.

I lead the Language and Artificial Intelligence Research (LAIR) group at Tennessee Tech University.

Deep learning fundamentals

What can neural networks compute, why do they train, and how can we tell what they've actually learned? This work ranges from proofs about the computational power of decoder-only transformers to tools like Torch Canvas that make architectures, initialization and signal propagation visible.

Selected papers

Torch CanvasInteractive demosCourse: Deep Learning (AI 4200/5200)

Large language models

How large language models represent uncertainty, reason, and learn, and how their behavior compares with human cognition: calibration and human-aligned uncertainty, memory effects, test-taking strategies, and alignment from feedback.

Selected papers

All 24 papers in this area are on the Publications page →

Seminar: Special Topics on LLMs (CSC 7970)LLM foundations reading list

Cherokee language revitalization

Language technology in support of Cherokee (Tsalagi) revitalization, built with and for the community. Drawing on what worked for Irish (Gaeilge), the goal is a living resource that supports future generations of Cherokee speakers, along with the methods low-resource, polysynthetic languages need, such as better tokenization.

CORE Institute Fellow: Indigenous Language Revitalization CORE Institute · 2027

Selected for a small cohort that will shape what responsible, community-centered AI looks like in Indigenous language revitalization, collaborating on concrete projects over the 2027 fellowship year.

Selected papers

In the news

Spiking neural networks Emerging

Brain-inspired networks that compute with discrete spikes over time. This newer line of work builds on earlier research into biologically extending artificial neural networks, and asks how spiking models can be trained and analyzed with the same rigor as conventional deep networks.

Early work

Tornado detection Emerging

Applying deep learning to severe-weather data to detect tornadoes more reliably and earlier, where better detection can directly save lives.

New work in progress; papers will appear here as they're published.

Funding

Join the group

I'm accepting master's and Ph.D. students. If you're curious and want to work on real problems with AI, email me or book a meeting.