Scalable Reconfigurable Architecture for Deep Learning

This research develops novel domain-specific processors to enable real-time processing over a pixel input stream close to sensors. The proposed AI-processors provide server-class AI computing capabilities with significantly lower power consumption. It can become a key building block for building real-time video analytic systems at the edge. The key feature of our proposed processors is direct stream processing for a diverse range of deep learning and video analytics algorithms without the need to storing the input pixel data. This is also very essential to mitigate the privacy and social concerns related to the deployment of video analytics in community areas. Additionally, it will also help to reduce unnecessary data movement throughout the memory hierarchy (keeping the data local), as well as to provide a more deterministic execution behavior.