Two papers accepted to NeurIPS 2026
Work on concept transfer and transparent reporting will appear at this year's conference.
VAIL · est. 2023
Our team builds models that see the world, and tools that explain how and why they work.
focus areas
Our research focus
We study perception, generalization, and explanation in modern AI systems. From robust recognition and segmentation to attribution, saliency, and concept-based explanations, we aim to make AI more reliable and interpretable.
computer-vision
Computer Vision
Detection, segmentation, multi-modal learning, and evaluation on real-world datasets.
interpretabilityInterpretability
Attributions, saliency maps, concept probes, and human-centered evaluation for explanations.
responsible-aiResponsible AI
Robustness, fairness, data quality, and transparent reporting for trustworthy systems.
3 selected
Selected work
interpretability
interpretability
Attribute
An open-source toolkit for generating and evaluating attribution maps across vision architectures, from CNNs to ViTs.
responsible-ai
responsible-ai
FairSight
A dataset auditing tool that surfaces demographic imbalance and labeling inconsistency before training begins.
computer-vision
computer-vision
ShiftBench
A benchmark suite for measuring detection and segmentation performance under real-world distribution shift.
recent
Recent publications
How Well Do Concepts Transfer Across Multi-Modal Tasks?
Yuki Tanabe, Mira Solheim · NeurIPS · 2026
Counterfactual Attribution Maps for Vision Transformers
Sana Idris, Mira Solheim · CVPR · 2026
Auditing Large-Scale Vision Datasets for Demographic Fairness
Priya Natarajan, Dev Anand · FAccT · 2025
A Benchmark for Failure-Mode Detection Under Distribution Shift
Dev Anand, Mira Solheim · ICCV · 2025
lab news
News
Attribute toolkit released
Our open-source attribution toolkit is now public, with support for CNNs and vision transformers.
Sana Idris defends PhD thesis
Congratulations to Sana on a successful defense of work on counterfactual attribution.
5 members
People
get in touch
Contact
We publish open research, release datasets and tools, and collaborate with academia and industry.
Contact the lab