Zoom: https://illinois.zoom.us/j/88904492060?pwd=RlB4SmZEbzg1OTJ6YjJ1WEhvTXdBdz09
Abstract:
Every day, a vast amount of data is generated, and edge systems—which refer to any computing agent beyond large-scale datacenter machines—play an increasingly crucial role in processing this deluge of data. Unlike conventional systems that are engineered with abundant resources, edge systems operate in real-world conditions, facing a multitude of unknown design-space trade-offs with limited resources that hinder their full-scale autonomy. This leads to isolated, time-consuming, and costly approaches to each challenge, often leading to suboptimal edge systems. To effectively navigate the constraints and unique multidimensional design space of edge systems, my research crafts innovative machine learning techniques and harnesses hardware-software synergy to establish roadmaps across the hardware-software stack for the next generation of edge systems.
In my talk, I first use the example of quadcopter drones to show how my research pioneers revealing the unique multidimensional design space of edge systems and suggests optimal points within this space, depending on the use case. By formulating the fundamental drone subsystems and introducing our open-source customizable drone, I explain how computations impact this design space. For instance, by exploring simultaneous localization and mapping (SLAM) implementations on various hardware platforms (CPU, GPU, FPGA, and ASIC), I demonstrate which approach is best suited for drone applications. The second part of my talk emphasizes the necessity of modern machine learning techniques, such as those utilizing deep neural networks, in comprehending complex raw data in edge systems and acting upon the outcomes. I show how my research empowers edge devices to overcome individual resource constraints through distributed computation among collaborating peer devices and proposes edge-aware neural networks by exploring hardware-software co-designs, algorithmic modifications, and system-level optimizations. In the end, I propose my plans for effectively handling data in exotic frontiers of edge systems with unique constraints, paving the way for innovative applications that will shape our future.
Bio:
Ramyad Hadidi is an applied machine learning researcher at Rain, where he leverages his expertise in edge computing to develop advanced artificial brains. Prior to joining Rain, Ramyad worked as a machine learning researcher at SK hynix, focusing on the efficient execution of cutting-edge computer vision workloads at the intersection of hardware, software, and CMOS image sensors (CIS). Ramyad earned his Ph.D. in Computer Science from the Georgia Institute of Technology in May 2021, under the guidance of Professor Hyesoon Kim. His thesis, titled "Deploying Deep Neural Networks in Edge with Distribution," explored the challenges and opportunities in edge computing and machine learning. Ramyad's research interests span a diverse range of fields, including computer architecture, robotics, edge computing, and machine learning. In addition to his dissertation research, Ramyad has made contributions to studies on processing-in-memory, GPU systems, and hardware accelerators for sparse problems. He believes that striking a balance between depth and breadth in research is key to discovering and tackling authentic research challenges. Webpage— https://ramyadhadidi.github.io/
Faculty Host: Josep Torellas
Meeting ID:889 0449 2060 ; Password: csillinois
