Towards Energy-Efficient Systems for Artificial Intelligence in the Future

Cognitive Robotics & Machine Learning powered by Artificial Intelligence (AI) are now playing significant roles in various domains. The rapid development of AI is propelled by three crucial components: large-scale data, AI algorithms, and computation circuits and systems. The circuits and systems provide fundamental computation capability for analyzing data and executing algorithms. Specific and heterogeneous circuits and systems are propping up current AI computation capability. However, with the volume of data becoming larger and larger, and the slowing down of Moore's Law, currently circuits and systems for AI is now facing great challenges in the future. From the data perspective, large-scale data are organized using sparse structures including graphs, networks, time series, and spiking signals. However, nowadays circuits and systems are highly structured, which are far away from analyzing and handling large-scale sparse data efficiently. From the algorithm perspective, collaborative intelligence becomes a promising way to surpassing the computation capability limitation of a single node, while currently the performance of collaborative intelligence algorithms is constrained by limited communication resources, complex data dependency, and lacking automation tools.To overcome these problems and provide energy-efficient circuits and systems to propel future AI, the structured sparse design and collaborative perception/decision methods will be introduced. The hardware-software co-design idea is introduced in the structured sparse design to map and process unstructured sparse data on current structured circuits and systems efficiently. While the variable center framework is adopted in the collaborative intelligence systems to realize collaborative perception/decision by multiple agents. All these designs will propel the development of AI in various domains in the future, including autonomous driving, recommendation systems, and etc.