Vector Researchers Showcase Machine Learning Applications at CVPR 2021
Researchers from the Vector Institute for Artificial Intelligence are presenting a diverse array of machine learning applications at the virtual Conference on Computer Vision and Pattern Recognition (CVPR) 2021. The showcased work spans multiple industries, including healthcare, security, and beauty technology. Notable projects include LOHO, a collaboration with Modiface that uses generative models to virtually alter hairstyles and colors in photos. In the realm of cybersecurity, researchers demonstrated vulnerabilities in machine learning systems through Data-Free Model Extraction, showing how models can be stolen without access to training data, particularly threatening on-device applications. They also proposed defenses via dataset inference techniques. Additionally, the team introduced AdvSim, an adversarial framework designed to generate safety-critical scenarios for self-driving vehicles by modifying LiDAR sensor data to identify potential autonomy failures. Another highlighted project, DatasetGAN, aims to create efficient labeled data factories with minimal human effort. These presentations underscore the institute's contribution to advancing computer vision and addressing practical challenges in AI deployment, safety, and security.
Wire timeline
Vector Researchers Showcase Machine Learning Applications at CVPR 2021
Researchers from the Vector Institute for Artificial Intelligence are presenting a diverse array of machine learning applications at the virtual Conference on Computer Vision and Pattern Recognition (CVPR) 2021. The showcased work spans multiple industries, including healthcare, security, and beauty technology. Notable projects include LOHO, a collaboration with Modiface that uses generative models to virtually alter hairstyles and colors in photos. In the realm of cybersecurity, researchers demonstrated vulnerabilities in machine learning systems through Data-Free Model Extraction, showing how models can be stolen without access to training data, particularly threatening on-device applications. They also proposed defenses via dataset inference techniques. Additionally, the team introduced AdvSim, an adversarial framework designed to generate safety-critical scenarios for self-driving vehicles by modifying LiDAR sensor data to identify potential autonomy failures. Another highlighted project, DatasetGAN, aims to create efficient labeled data factories with minimal human effort. These presentations underscore the institute's contribution to advancing computer vision and addressing practical challenges in AI deployment, safety, and security.
Vector Institute for Artificial Intelligence