Autoregressive Simulation and Synthetic Data for Geometric Vision
Details
Autoregressive Simulation and Synthetic Data for Geometric Vision
Agenda:
18:00-18:15: Gathering and mingling
18:15-19:00: Autoregressive Simulation of Many-body Quantum Systems, by Or Sharir
19:00-19:45: Synthetic Data for Geometric Computer Vision, by Matan Sela
Autoregressive Simulation of Many-body Quantum Systems
The theoretical understanding and modeling of interacting many-body quantum matter represents an outstanding challenge . At the heart of several problems in condensed matter, chemistry, nuclear matter, and more lies the intrinsic difficulty of fully representing the many-body wave-function, in principle needed to exactly solve Schrödinger’s equation, which grows exponentially with the number of particles.
In this talk, I will present our contributions in this field that were recently accepted to Physical Review Letters, the most prestigious journal in all of physics. First, we argue that deep convolutional networks are a better alternative to RBMs, by proving that they can represent highly entangled quantum states far more efficiently. Second, we propose a specialized architecture for modeling wave-functions to overcome the optimization limitations, inspired by neural autoregressive models. We demonstrate that our method allows for the first time to employ large-scale networks with millions of parameters on this problem, which leads to state-of-the-art accuracy while using significantly less computational resources overall relative to competing methods.
Or Sharir is concluding his Ph.D. in Computer Science at the Hebrew University of Jerusalem under the supervision of Prof. Amnon Shashua. His research ranges from the theoretical analysis of Deep Learning methods to the application of such methods to other fields, such as Language Understanding and Theoretical Physics. Or also holds a B.Sc. degree in Physics, Mathematics, and Computer Science from the Hebrew University.
Synthetic Data for Geometric Computer Vision
The generalization capability of deeply architectured models is the main trigger behind the rise of artificial intelligence in recent years. Although these models are irrationally useful for solving supervised learning problems, they are data hungry. Acquisition of accurately annotated real data poses a major barrier for most model training endeavors. An alternative approach is to synthesize data by imitating the process by which real data is generated. In this talk, I will present five of my publications enclosing different solutions to complex computer vision problems with geometric flavor. In between, I will discuss the strengths and limitations of various methods to synthesize data including adversarial training.
Dr. Matan Sela is a co-founder and CTO at Mirrori. He obtained his Ph.D. at the Faculty of Computer Science in the Technion in 2018. His research focused on combining deep learning with computer graphics for solving geometric computer vision problems. Matan earned B.Sc. and M.Sc. in Electrical Engineering from the Technion, both with honors.