Jason Corso

Jason Corso is a computer-vision researcher and co-founder of Voxel51. With Jacob Marks and Brent A. Griffin, he developed a framework for measuring classification-dataset difficulty and detecting potential label mistakes. As of September 2026, he is Voxel51’s chief science officer and a professor of Robotics and Electrical Engineering and Computer Science at the University of Michigan, where he holds the Toyota Professorship of Artificial Intelligence. His research publications use the name Jason J. Corso.[3][2][4][5]

Corso, Marks, and Griffin’s Class-wise Autoencoders Measure Classification Difficulty And Detect Label Mistakes, submitted on December 3, 2024, analyzes datasets through reconstruction error ratios (RERs) between autoencoders trained on individual classes. The framework characterizes difficulty at the sample, class, and whole-dataset levels. It separates contributions from finite sample size and from Bayes error and decision-boundary complexity.[4]

Across 19 visual datasets, the authors reported a strong correlation between their dataset-difficulty probe and the error rates of state-of-the-art classification models. They also interpreted sample-level difficulty as a label-mistakenness score, reporting state-of-the-art mislabel detection on hard datasets under symmetric and asymmetric label noise. These findings concern the study’s tested datasets and noise conditions.[4]

The joint work includes a public repository for computing RERs and reproducing the experiments. Its FiftyOne integration stores a mistakenness score and a flag for each sample, allowing users to filter and rank potential label mistakes for inspection. A flag identifies a suspected error rather than confirming that a label is wrong. In his February 4, 2025 technical explanation, Corso explicitly credited the collaboration: “This blog gently introduces technical work done by Jacob Marks, Brent Griffin, and myself.”[5][6]

Corso appeared in Latent Space’s December 30, 2023 NeurIPS startup recap, which introduced him as Voxel51’s CEO and linked the open-source FiftyOne project. That historical role differs from his current chief science officer position.[1][3]

Before the dataset-difficulty work, Corso’s research included video understanding. He and Sreemanananth Sadanand jointly authored Action Bank: A High-Level Representation of Activity in Video, with both authors affiliated with Computer Science and Engineering at SUNY at Buffalo. This earlier collaboration supplies context for his work in computer vision alongside his later research on classification datasets.[7][4]