Deep Neural Networks (DNNs) have emerged as the promising machine learning technology for a number of inferencing/classification problems including object detection, face recognition, gesture recognition, voice command recognition, sound signature detection, anomaly detection etc. For always-on sensor node applications, where the events of interest occur infrequently, there is a need to detect events locally at the end-node as against the cloud, so as to reduce network traffic, operate with no or unreliable connectivity, enhance privacy (only event information sent to the cloud as against raw data) and reduce response-time. DNNs however are compute and data intensive, so implementing them on end-nodes under aggressive cost, power and latency constraints is a big challenge. In this talk we address the challenge of cost and power optimized DNN realizations while meeting the accuracy and latency requirements of the target application. Our analysis shows that 2-3 orders of magnitude reduction in die size * power  metric is needed to make ML at the Edge a reality. We present multiple techniques spanning system, algorithm, architecture, design, circuit and technology level optimizations to address this challenge. We also highlight open problems/areas of further research to keep pushing the envelope.