Applications such IoT, the Industrial Internet, robotics and autonomy, as well as wearable devices are heralding the onset of the true world of ubiquitous connected computing. There are many reasons why it makes perfect sense to make these distributed edge devices “intelligent” as well: communication bandwidth, latency, robustness, security and privacy are just a few of them. Yet, given the limited power and volume budget, machine-learning techniques and architectures that are successsful in the cloud may not be directly amenable to the edge. Especially, the need for agility, continuous adaptation and customization requires learning approaches that are fragile, and may operate on “little” data rather than the “big data” approach, used for instance in deep nets. Getting the orders of magnitude in energy-efficiency improvement leads to the adoption of alternative approaches towards computation that require innovation across the implementation hierarchy – from technology, devices, circuits, architectures and systems.