Oncel Tuzel

Also known as Cuneyt O. Tuzel

Oncel Tuzel is a machine learning researcher whose computer vision and generative-model work spans robot vision at Mitsubishi Electric Research Laboratories and products at Apple. He earned a Rutgers PhD under Peter Meer in 2008, and his CVPR 2007 paper on classification on Riemannian manifolds was a Best Paper runner-up. At MERL, his research earned the 2014 R&D 100 Award for the MELFA 3D Vision System, and in 2016 he and Ming-Yu Liu introduced coupled generative adversarial networks. At Apple, his team's 2017 paper on learning from simulated and unsupervised images won a CVPR Best Paper Award, and he coauthored VoxelNet in 2017 and the GSM-Symbolic benchmark in 2024. He is a Distinguished Scientist on Apple's Machine Learning Research team, where he works on efficient machine learning, controllable generative AI and data-centric ML.

Now at: Apple
Badges: Bridge

Timeline

2007

Human detection on Riemannian manifolds

At CVPR 2007, he presents 'Human Detection via Classification on Riemannian Manifolds' with Fatih Porikli and Peter Meer. The paper treats covariance matrices as points on a Riemannian manifold to detect humans in still images, and it takes the Best Paper runner-up award at the conference.
2008

Rutgers PhD under Peter Meer

He receives his PhD in computer science from Rutgers in 2008, supervised by Peter Meer, with a thesis titled 'Learning on Riemannian Manifolds for Interpretation of Visual Environments'. His journal paper 'Pedestrian Detection via Classification on Riemannian Manifolds', written with Fatih Porikli and Peter Meer, appears in IEEE Transactions on Pattern Analysis and Machine Intelligence in October 2008.
2014–2015

Bin-picking vision at MERL

At Mitsubishi Electric Research Laboratories in Cambridge, his research earns the 2014 R&D 100 Award for the MELFA 3D Vision System. In December 2015, he gives a keynote talk at the International Symposium on Visual Computing in Las Vegas, titled 'Machine vision for robotic bin-picking: Sensors and algorithms'.
2016

Coupled generative adversarial networks

With Ming-Yu Liu, he proposes coupled generative adversarial networks (CoGAN), which learn a joint distribution of multi-domain images without any tuple of corresponding images. The paper appears at NIPS 2016 in Barcelona.
2017

A CVPR best paper at Apple

His team at Apple wins the CVPR 2017 Best Paper Award for 'Learning from Simulated and Unsupervised Images through Adversarial Training', written with Ashish Shrivastava, Tomas Pfister, Josh Susskind, Wenda Wang and Russ Webb. The paper proposes Simulated+Unsupervised learning, which uses an adversarial network to improve the realism of simulator images with unlabeled real data, and it reaches state-of-the-art results on the MPIIGaze dataset without labeled real data.
November 2017

VoxelNet for 3D detection

In November 2017, he and Yin Zhou post VoxelNet on arXiv, a single-stage, end-to-end trainable network that divides a LiDAR point cloud into 3D voxels and learns a feature representation for each voxel. VoxelNet outperforms earlier LiDAR-based 3D detectors on the KITTI car detection benchmark.
October 2024

The GSM-Symbolic benchmark

In October 2024, he coauthors GSM-Symbolic with Iman Mirzadeh, Keivan Alizadeh, Hooman Shahrokhi, Samy Bengio and Mehrdad Farajtabar. The benchmark builds math questions from symbolic templates, and the study finds that the performance of all tested models declines when only the numerical values in a question change.
As of October 2026

Distinguished Scientist at Apple

He is a Distinguished Scientist on the Machine Learning Research team at Apple, working on efficient ML algorithms, controllable generative AI and data-centric ML. His work has contributed to Siri Invocation, Siri ASR, Scribble, Keyboard Intelligence, Face Recognition, VisionKit, Apple Maps, Apple Photos and Vision Pro. In 2026, he coauthors ICLR papers including 'LiTo: Surface Light Field Tokenization' and 'Pretraining with Hierarchical Memories: Separating Long-Tail and Common Knowledge'.