Research Topic 1: Computing Systems for Linear Algebra
The advancement of computing systems, arguably the most pivotal enabler of human discovery, faces fundamental constraints from physical limitations and finite energy resources. Transistor density scaling, the primary methodology for increasing computing power for decades, is approaching saturation. Concurrently, the energy consumption of computing systems is outstripping energy production, contributing to carbon emissions and exacerbating climate change concerns. An effective solution involves the intelligent utilization of transistors through the integration of application-specific accelerators alongside existing generic processors, resulting in heterogeneous computing paradigms. However, a significant challenge arises from the tedious manual restructuring required to adapt current applications and algorithms for these diverse resources. Although automated systems, such as compilers, are crucial for facilitating application redevelopment for heterogeneous processors, they frequently lag behind the rapid proliferation of novel accelerators and the communication bottlenecks inherent in these architectures. This project seeks to develop integrated software and hardware systems designed to accelerate scientific and machine learning applications on future heterogeneous architectures, thereby achieving high-performance and resilient systems with sustainable energy consumption.
Research Topic 2: Compiler for Cyber-Physical Systems
This project aims to develop scalable AI and optimization software systems specifically for resource-constrained computing environments, such as edge devices and industrial controllers within cyber-physical systems (CPS). While optimization and AI methods, particularly machine learning (ML), are crucial for applications like drone navigation, autonomous driving, medical signal processing, and power flow optimization, they often struggle to meet real-time requirements on devices with limited memory, power, and computational capabilities. Current state-of-the-art frameworks are typically optimized for high-performance processors equipped with GPUs and specialized accelerators, which are rarely found in CPS devices. We propose to develop real-time software for CPS devices that is inherently aware of power consumption and physical constraints..
Research Topic 3: Big Data Analytics and Machine Learning
This research project addresses the formidable challenges posed by the exponential growth of big data, particularly within critical domains like healthcare and complex scientific simulations. Our core objective is to advance knowledge discovery by developing novel AI models capable of extracting profound insights from these massive datasets, strategically leveraging the strengths of existing state-of-the-art models rather than starting from scratch. Beyond model creation, a significant thrust of our work involves analyzing big data to identify intrinsic patterns that enable substantial reductions in their memory footprint, thereby mitigating the strain on storage and processing resources. Concurrently, we are developing sophisticated techniques to reduce the memory size of the resulting AI models themselves without compromising their accuracy or predictive power, a crucial step for deploying advanced AI solutions efficiently in diverse, potentially resource-constrained, real-world applications.