Learning Semantic Image Representations at a Large Scale
Yangqing Jia · UC Berkeley · 2014
I present my work towards learning a better computer vision system that learns and generalizes object categories better, and behaves in ways closer to what human behave. Specifically, I focus on two key components of such a system: learning better features, and revisiting existing problem statements. For the first component, I propose and analyze novel receptive field learning and dictionary learning methods, mathematically justified by the \ ystrom sampling theory, that learn more compact and effective features for object recognition tasks. For the second component, I propose to combine otherwise independently developed computer vision and cognitive science studies, and present the first large-scale system that allows computers to learn and generalize closer to what a human learner will do. I also provide a large-scale human behavior database, which will hopefully enable further research along this research direction.Following the recent success of convolutional neural networks, I present and release a well-engineered framework for general deep learning research, and provide an extensive analysis on the generality of deep features learned from the state-of-the-art CNN pipeline: whether they serve as a general-purpose visual descriptor that could be adopted in various applications, and future research directions made possible by such general features.