Machine Intelligence
Machine intelligence is simultaneously one of the most consequential and least understood areas of modern computing. INSTAR's research examines how AI and ML systems learn, generalize, and fail — with equal weight given to capability and to safety, reliability, and responsible deployment. Our interdisciplinary approach draws from statistics, cognitive science, mathematics, and domain sciences to ask not only what models can do, but what they should do and when they should not be trusted.
Natural Language Processing & Language Models
INSTAR examines how large language models represent and manipulate knowledge — and, critically, where they fail to do so reliably. Research interests include retrieval-augmented generation, multilingual understanding, factual grounding, and the evaluation methodologies needed to assess model reliability in high-stakes scientific and operational contexts. The goal is not to chase benchmark performance but to understand what linguistic capability actually means for real applications — scientific synthesis, structured information extraction, and research assistance where errors have downstream consequences.
Computer Vision & Perception
Computer Vision & Perception
Visual data — medical images, remote sensing imagery, scientific microscopy, and video streams — carries information that is difficult to formalize but essential to act on. INSTAR investigates visual recognition architectures for object detection, semantic segmentation, and 3D scene reconstruction, with emphasis on domains where training data is scarce, labels are expensive, and errors are costly. We are particularly interested in uncertainty-aware perception: systems that can signal when their outputs should not be trusted, which is often more valuable than marginal accuracy gains on standard benchmarks.
Reinforcement Learning & Autonomous Systems
Reinforcement learning offers a principled framework for sequential decision-making under uncertainty, but its practical deployment demands far more than benchmark scores. INSTAR studies sample-efficient learning, safe exploration, and the robustness of learned policies under distribution shift — questions that determine whether RL-derived behavior is suitable for real laboratory automation, resource-management systems, and multi-agent coordination scenarios. The INSTAR Consortium Postdoctoral Fellowship engages early-career researchers from AI, physics, mathematics, and cognitive science; apply at /fellowship/.
OUR PARTNERS
