Enabling Deep Intelligence on Embedded Systems

As deep learning for resource-constrained systems become more popular, we see an increased number of intelligent embedded systems such as IoT devices, robots, autonomous vehicles, and the plethora of portable, wearable, and mobile devices that are feature-packed with a wide variety of machine learning tasks. However, the performance of DNNs (deep neural networks) running on an embedded system is significantly limited by the platform's CPU, memory, and battery-size; and their scope is limited to simplistic inference tasks only. This dissertation proposes on-device deep learning algorithms and supporting hardware designs, enabling embedded systems to efficiently perform deep intelligent tasks (i.e., deep neural networks) that are high-memory-footprint, compute-intensive, and energy-hungry beyond their limited computing resources. We name such on-device deep intelligence on embedded systems as Embedded Deep Intelligence. Specifically, we introduce resource-aware learning strategies devised to overcome the four fundamental constraints of embedded systems imposed on the way towards Embedded Deep Intelligence, i.e., in-memory multitask learning via introducing the concept of Neural Weight Virtualization, adaptive real-time learning via introducing the concept of SubFlow, opportunistic accelerated learning via introducing the concept of Neuro.ZERO, and energy-aware intermittent learning, which tackles the problems of the small size of memory, dynamic timing constraint, low-computing capability, and limited energy, respectively. Once deployed in the field with the proposed resource-aware learning strategies, embedded systems are not only able to perform deep inference tasks on sensor data but also update and re-train their learning models at run-time without requiring any help from any external system. Such an on-device learning capability of Embedded Deep Intelligence makes an embedded intelligent system real-time, privacy-aware, secure, autonomous, untethered, responsive, and adaptive without concern for its limited resources.Doctor of Philosoph