"Computational Sensors for Energy-efficient Deep Learning"
Computer vision and deep learning has had tremendous success in recent years, driving applications from object detection, recognition/tracking, segmentation, and other robotics applications. But the energy costs of image sensing and processing can be prohibitive for embedded vision, limiting battery life and deployment in the field. In this talk, I'll present our research on using specialized image sensors to capture features for computer vision directly, and that are energy-efficient. I'll first discuss a way to optically compute the first layers of convolutional neural networks using custom CMOS diffractive image sensors called Angle Sensitive Pixels. Then I will present research on the specialized image signal processing (ISP) pipelines for computer vision. By modifying the abstraction and design boundaries between computational sensors and processing, we can envision new possibilities for energy-efficient visual computing in the future. This talk will be accessible for a general audience and will introduce concepts from sensors, computer vision, and machine learning as needed.