CS Talk
Chunwei Liu, MIT
Refreshments from Koffee Katering will be availalble.
Host: Holly Rushmeier
Title: Adaptive Data Systems for All the World’s Bytes
Abstract:
Modern enterprises produce massive volumes of data, while important applications ranging from conventional data analytics to AI-powered applications like large language models demand extensive computational resources. At the same time, Moore’s Law is reaching its limits. In this talk, I will present my research on two adaptive data systems that address these data and computation challenges through co-design and decomposition. Both systems separate high-level task specifications from low-level implementation, allowing users to focus on what to compute while the system determines how to achieve an optimized implementation.
First, I will introduce AdaEdge, an adaptive data compression framework designed for massive data volumes beyond storage or bandwidth capacity. By co-designing compression and data processing, AdaEdge automatically selects compression methods based on user preferences, workloads, and system constraints, thereby increasing efficiency without compromising performance in downstream tasks. Next, I will present Palimpzest, an optimized data system for complex AI-powered pipelines. It offers a declarative interface for processing unstructured data, enabling database-style optimizations that significantly reduce inference costs while preserving output quality.
Bio:
Chunwei Liu is a Postdoctoral Associate in the Data Systems Group at MIT CSAIL. His research interests span compound AI systems, database systems, cloud/edge computing, and database benchmarking. His current focus is on optimizing data systems for both conventional data analytics and emerging AI-powered pipelines. He also led a cloud database benchmarking effort in collaboration with major industry leaders, including Microsoft, Amazon, Intel, Meta, and Google. Chunwei’s research has been published in top-tier data management (SIGMOD, VLDB, ICDE & CIDR) and AI (ICLR, EMNLP & NAACL) venues. His work was recognized as the Best Paper Runner-Up at VLDB 2023 and invited to the “Best of VLDB” special issue of the VLDB Journal. Chunwei received his Ph.D. from the University of Chicago, where he was part of the ChiData group.