1:50 PM - 2:10 PM
[ACG43-02] Deep Learning that Ensures Laws of Physics and Other Constraints and Its Applications
★Invited Papers
Keywords:artificial intelligence, physics simulation, geometric structure, prior knowledge
Machine learning and artificial intelligence are attracting much attentions, mainly thanks to the flexibility of deep learning. When feeding big data to deep learning, it learns meaningful representations and builds a sophisticated model of decision making automatically. However, the flexibility is not the sole requirements. For example, deep learning is expected to learn physical phenomena and to accelerate the physical simulations. Then, deep learning has to reproduce the underlying geometric structure that admits laws of physics rather than the superficial dynamics. Indeed, most of deep learning methods gain outstanding performances by being associated with certain symmetries and constraints based on prior knowledge, e.g., a convolutional neural network is designed to be shift-invariant. The same can be said of probabilistic modeling and reinforcement learning. This talk introduces such deep learning based on prior knowledge.