Abstract: This talk will cover our recent and ongoing work on developing Computer Vision and Machine Learning methods that when implemented on a Pixel Processor Array (PPA), enable inference and learning at the image plane. A PPA allows this due to its parallel nature of computation and its ability to process images at source. The result is that a sensor-processor array can make inference and decide what to send down the visual pipeline and importantly without having to send images out if these are not relevant. This has important consequences for IoT devices, mobile systems and Robotics where computational and energy efficiency is crucial and where there is concern of privacy due to capturing more than what the tasks actually require. In this talk we will focus on work developed within the EPSRC AGILE project between the universities of Manchester, who developed the PPA SCAMP architecture, and our work at University of Bristol where we have developed parallel Computer Vision and CNN methods for a variety of tasks in classification and navigation. Specifically we will cover methods for implementing CNNs on PPAs, the way they operate, training approaches, efficient implementations to reach tens of thousands of fps and their applications.