what we study
Research
We do two things: build systems that perceive, and build methods that make their reasoning legible. The second is what makes the first usable, because a model whose decisions nobody can inspect is one nobody can debug, audit, or safely deploy.
computer-vision
Computer Vision
Recognition, segmentation, generation, and multimodal systems, with a focus on data and evaluation.
interpretabilityInterpretability
Attribution, concept probes, and mechanistic analysis of what models represent internally.
ai-safetyAI Safety
Failure modes, robustness under shift, and evaluation that catches problems before deployment.