Understanding the Visual World Through Naturally Supervised Code

Explore the intersection of symbolic structure and neural networks in visual understanding through this lecture by Jiajun Wu from Stanford. Delve into how symbolic code can be learned from natural supervision, including pixels, objects, and language. Examine the complementary roles of symbolic programs and neural networks in capturing high-level structure and extracting complex features from visual and language data. Discover methods for inferring, representing, and utilizing symbolic structure from raw data without compromising neural network expressiveness. Learn about neuro-symbolic approaches for scene synthesis, regular intrinsics inference, and grounded visual concept learning. Gain insights into the data efficiency and generalization capabilities of symbolic programs compared to deep neural networks in visual understanding tasks.

Understanding the Visual World Through Naturally Supervised Code