LEARNING-TO-VR.GITHUB.IO

Visual Recognition, as a key research field in artificial intelligence, has enabled rapid progress over the past years. Learning methodologies from various perspectives have achieved remarkable performance in tasks including image recognition, image generation, visual question-answering, visual navigation, video analytics, robotics, etc. In spite of these successes, visual recognition still faces many challenges. Humans can learn new concepts with very little supervision while the most powerful deep learning techniques might fail. This motivates the machine learning direction of zero-/few-shot learning as well as meta-learning. Besides, other obstacles such as generalization problem caused by learning with multiple domains are deserved to be investigated. It is fairly difficult to build feature representations that are transferrable between source domains and target domain. It gives rise to explorations on efficient data usages and training procedures during modeling. In addition,..