Joint Training of a Neural Network and a Structured Model for Computer Vision

Identifying objects and telling where they are in real world images is one of the most important problems in Artificial Intelligence. The problem is challenging due to: occluded objects, varying object viewpoints and object deformations. This makes the vision problem extremely difficult and cannot be efficiently solved without learning. This thesis explores hybrid systems that combine a neural network as a trainable feature extractor and structured models that capture high level information such as object parts. The resulting models combine the strengths of the two approaches: a deep neural network which provides a powerful non-linear feature transformation and a high level structured model which integrates domain-specific knowledge. We develop discriminative training algorithms to jointly optimize these entire models end-to-end. First, we proposed a unified model which combines a deep neural network with a latent topic model for image classification. The hybrid model is shown to outperform models based solely on neural networks or topic model alone. Next, we investigate techniques for training a neural network system, introducing an effective way of regularizing the network called DropConnect. DropConnect allows us to train large models while avoiding over-fitting. This yields state-of-the-art results on a variety of standard benchmarks for image classification. Third, we worked on object detection for PASCAL challenge. We improved the deformable parts model and proposed a new non-maximal suppression algorithm. This system was the joint winner of the 2011 challenge. Finally, we develop a new hybrid model which integrates a deep network, deformable parts model and nonmaximal suppression. Joint training of our hybrid model shows clear advantage over train each component individually, and achieving competitive result on standard benchmarks.