Self-supervised learning for images, video, and 3D - Ishan Misra, Meta | GHOST Day: AMLC 2022 скачать в хорошем качестве

Self-supervised learning for images, video, and 3D - Ishan Misra, Meta | GHOST Day: AMLC 2022

GHOST Day: AMLC 2022 About the speaker: Ishan Misra finished his Ph.D. at the Robotics Institute at Carnegie Mellon University in 2018. He has since then been working as a Research Scientist at Facebook AI Research (FAIR). His main research interests are Computer Vision and Unsupervised Learning, having published multiple research papers on Self-Supervised Learning and Visual Representation, together with prominent researchers like Yann LeCun and Martial Hebert. Ishan's works have won multiple awards such as the best paper award at WACV 2014 and best paper nomination at CVPR 2021. Ishan was also a guest on the Lex Fridman Podcast and ML Street Talk. Abstract: Supervised learning has been the primary success story in computer vision. Pretraining on large, labeled data leads to highly transferable feature representations. In this talk, I will present self-supervised methods we developed at FAIR that can learn representations that surpass or match the quality of supervised pretrained methods. All these methods are based on the simple principle of learning representations that are invariant to visual transforms. This simple principle leads to powerful methods that can be easily applied to image, video, and 3D data, and can leverage large amounts of unlabeled data. The resulting self-supervised models can be used via transfer learning to create state-of-the-art object detectors, action recognition and 3D recognition models. Self-supervised pretraining leads to more robust representations and can also help with `tail’ classes in recognition. Beyond transfer learning, I will show how self-supervised methods can discover objects - discover pixels that can group together by using just image or audio signals.