CAREER: Efficient, Dynamic, Robust, and On-Device Continual Deep Learning with Non-Volatile Memory based In-Memory Computing System

Over past decades, there have existed grand challenges in developing high performance and energy-efficient computing solutions for big-data processing. Meanwhile, owing to the boom in artificial intelligence (AI), especially Deep Neural Networks (DNNs), such big-data processing requires efficient, intelligent, fast, dynamic, robust, and on-device adaptive cognitive computing. However, those requirements are not sufficiently satisfied by existing computing solutions due to the well-known power wall in silicon-based semiconductor devices, the memory wall in traditional Von-Neuman computing architectures, and computation-/memory-intensive DNN computing algorithms. This project aims to foster a systematic breakthrough in developing AI-in-Memory computing systems, through collaboratively developing a hybrid in-memory computing (IMC) hardware platform integrating the benefits of emerging non-volatile resistive memory (RRAM) and Static Random Access Memory (SRAM) technologies, as well as incorporating IMC-aware deep-learning algorithm innovations. The overarching goal of this project is to design, implement, and experimentally validate a new hybrid in-memory computing system that is collaboratively optimized for energy efficiency, inference accuracy, spatiotemporal dynamics, robustness, and on-device learning, which will greatly advance AI-based big-data processing fields such as computer vision, autonomous driving, robotics, etc. The research will also be extended into an educational platform, providing a user-friendly learning framework, and will serve the educational objectives for K-12 students, undergraduate, and graduate students. This project will advance knowledge and produce scientific principles and tools for a new paradigm of AI-in-Memory computing featuring significant improvements in energy efficiency, speed, dynamics, robustness, and on-device learning capability. This cross-layer project spans from device, circuit, and architecture to DNN algorithm exploration. First, a hybrid RRAM-SRAM based in-memory computing chip will be designed, optimized, and fabricated. Second, based on this new computing platform, the on-device spatiotemporal dynamic neural network structure will be developed to provide an enhanced run-time computing profile (latency, resource allocation, working load, power budget, etc.), as well as improve the robustness of the system against hardware intrinsic and adversarial noise injection. Then, efficient on-device learning methodologies with the developed computing platform will be investigated. In the last thrust, an end-to-end DNN training, optimization, mapping, and evaluation CAD tool will be developed that integrates the developed hardware platform and algorithm innovations, for optimizing the software and hardware co-designs to achieve the user-defined multi-objectives in latency, energy efficiency, dynamics, accuracy, robustness, on-device adaption, etc. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.