Ongoing Research

Practical ML for Edge Devices

Collaborators: Microsoft DeepSpeed, Palantir, POSTECH, Yonsei University

The demand for realistic machine learning systems tailored for edge devices has grown in recent years. A number of learning paradigms well-suited for edge devices have emerged, such as Continual Learning and Federated Learning. However, the adoption of such paradigms has been difficult as prior works do not adequately address the growing demand for system efficiency in practical edge environments.

In response, our team has focused on system solutions for cost-effective on-device learning. We build systems that optimize the energy-accuracy trade-offs and adapt to dynamic resource states of the target devices.

Related Publications: Miro, CarM

System Solutions for Large-scale ML 

Collaborators: ETRI, Meta, Microsoft Research Asia, Samsung Research, Ajou University, KAIST, Northeastern University

ML applications are increasingly requiring greater computing resources with the rapid growth in size of the state-of-art DNNs (e.g. LLMs). Many studies and productions have been released to address this problem. 

We strive to provide system solutions for different phases of DNN workloads such as efficient model parallelization for GPU clusters, improving GPU utilization, efficient data loading and preprocessing for DNN.

Related Publications: EnvPipeSibyllaHUVMZico, Philly, Tiresias

Fast and Scalable Big Data Analytics

Collaborators: Amazon, Microsoft Research Cambridge, Microsoft Research Redmond, Samsung Research, SNU, UIUC

Big data analytics plays a crucial role in today's data-driven world, offering the potential for improved decision-making, operational efficiency, and innovation across a wide range of domains.

We aim to build fast and scalable systems for real-time big data analytics at cloud/IoT scale that enable system operators to promptly troubleshoot system anomalies, improve the performance and reliability of their services, and optimize the performance tailored to specific workload characteristics.

Related Publications: Blaze, Sponge, SWAN, Jarvis, AOMG, StreamBox-HBM

System Solutions for Improving Food Production in the Field

Several studies emphasize the need for a substantial increase in global food production by 2050, making it crucial to use innovations in computer science to address various crop and livestock management issues for the next generation. However, farms are not typically equipped with ample and stable power supply and network connectivity, and traditional food production processes rarely harness the potential for intelligent decision-making driven by data generated from farm fields.

We aim to bring our expertise in big data analytics and machine learning to address global food production challenges. Our goal is to develop an end-to-end system stack covering data collection, ingestion, and processing, while efficiently utilizing extremely limited resources on farms.