Talk Abstract:  How should we represent 3D shapes in modern learning systems? While representations like meshes, voxels, 3D Gaussians, density and distance fields excel in computer graphics and physics simulations, one would argue that for machine learning, a good representation should be compact, differentiable, and easy to obtain at scale. In this talk, I will present our recent work on a novel 3D shape representation that satisfies these three criteria. Our representation is purely data-driven yet allows estimations of important geometric properties such as surface normals and deformation fields. Our representation shows strong performance in all machine learning tasks we tested, including single-image to 3D, point cloud to image and text alignment, and neural rendering.

 

Speaker Biography:  Rick Chang a senior research scientist at Apple. He has developed techniques for various topics, including neural rendering, AR/VR displays, language modeling, text-to-speech, and handwriting synthesis. His contributions have been incorporated into products like Siri's speech recognition, Scribble handwriting recognition, and QuickPath keyboard.
Before joining Apple in 2020, Rick Chang received his PhD from Carnegie Mellon University, where he built a projector that shows 2000 frames per second, a VR display that shows 3D even for users with one eye and solved various inverse problems. He received M.S and B.S degrees from National Taiwan University.