Deep learning algorithm optimization
Our goal is to minimize computation required for deep neural networks. We focus on vision tasks, e.g., object detection and segmentation. Our approach is to exploit redundancy in input data to the neural network. Currently, we focus on video data since video processing has potential of significant performance improvement by exploiting inter-frame correlation. Our optimizations, covering reinforcement learning, generative adversarial network, etc., judiciously take advantage of high-level correlation information, e.g., object class in previous frames.