Presentation
A Journey to SW/HW Co-design in Machine Learning: Fundamental, Advancement, and Application
TimeMonday, July 11th10:30am - 12pm PDT
Location3004, Level 3
Event Type
Tutorial
AI
EDA
DescriptionThe rapid evolution of machine learning (ML) has been enabling the integration of intelligence into various applications.
In today's ML-empowered applications, maximizing accuracy is no longer the only design objective; instead, the demands on hardware efficiency (e.g., latency, power) is sharply increasing. Driven by different objectives, two individual research threads, i.e., ML algorithm design and hardware acceleration, are being intensively investigated. However, to achieve the best accuracy-efficiency trade-off, two threads must be jointly studied and eventually merge into one journey.
In this tutorial, we aim to guide the interested audiences to the walk through the exhilarating journey towards efficient software/hardware co-design in ML-empowered systems, as well as its vast interests in a wide range of applications. Given the complexity of the topic and rich background knowledge required, we divide the tutorial into three talks, progressively leading to the goal from the fundamentals to advanced techniques, and to applications. The three talks will also cover different aspects of applications: autonomous systems, medical and drug discovery, and on-device augmented reality (AR).
Talk I aims to introduce the fundamentals of efficient ML acceleration design, as well as automated ML algorithm design, i.e., neural architecture search (NAS), which establishes the foundation for more advanced co-design techniques and applications. Talk I will also briefly introduce the co-design application in power-efficient autonomous systems. Talk II aims to discuss co-design technique advancements such as systematic co-design frameworks on top of Talk I, as well as real-world applications of co-design techniques with unique challenges and opportunities, especially fairness and privacy. Talk III specifically targets an emerging application from industry, augmented reality (AR), which has a sharply rising demand for sw/hw co-design for both vision and speech tasks.
In today's ML-empowered applications, maximizing accuracy is no longer the only design objective; instead, the demands on hardware efficiency (e.g., latency, power) is sharply increasing. Driven by different objectives, two individual research threads, i.e., ML algorithm design and hardware acceleration, are being intensively investigated. However, to achieve the best accuracy-efficiency trade-off, two threads must be jointly studied and eventually merge into one journey.
In this tutorial, we aim to guide the interested audiences to the walk through the exhilarating journey towards efficient software/hardware co-design in ML-empowered systems, as well as its vast interests in a wide range of applications. Given the complexity of the topic and rich background knowledge required, we divide the tutorial into three talks, progressively leading to the goal from the fundamentals to advanced techniques, and to applications. The three talks will also cover different aspects of applications: autonomous systems, medical and drug discovery, and on-device augmented reality (AR).
Talk I aims to introduce the fundamentals of efficient ML acceleration design, as well as automated ML algorithm design, i.e., neural architecture search (NAS), which establishes the foundation for more advanced co-design techniques and applications. Talk I will also briefly introduce the co-design application in power-efficient autonomous systems. Talk II aims to discuss co-design technique advancements such as systematic co-design frameworks on top of Talk I, as well as real-world applications of co-design techniques with unique challenges and opportunities, especially fairness and privacy. Talk III specifically targets an emerging application from industry, augmented reality (AR), which has a sharply rising demand for sw/hw co-design for both vision and speech tasks.