Curiosity-driven AI Systems Research


After more than 15 years experience in conducting advanced AI systems research and developing various enterprise AI solutions at IBM Research, I became a firm believer of conducting research through building functional AI systems and solutions that should be meaningful to the society and/or challenge the status quo.

Nowadays, one of the biggest challenges facing our society is the high cost, long cycle of developing innovative AI solutions to address complex societal issues, such as sustainability, education, accelerated scientific discovery, and fast response to address global crisis (such as the recent pandemics). I call it "Earthly AI" systems and solutions research (please take a look at the TEDxUIUC talk, "The Flashy and Earthly Sides of AI", that I gave togheter with Prof. Wen-mei Hwu on this subject).

One of my main research goals is to develop forward-looking applied AI technologies and solutions to improve the productivity of AI development and lower the cost of scalable AI deployment in multiple application domains.

In particular, I am adopting a curiosity-driven approach to identify the impactful research problems to solve. For example, I am interested in

  • building novel applied AI solutions and systems involving emerging technologies (such as IoT, AR/VR, edge clouds, and quantum computers) to address critical issues that matter the most to our society and humanities,

  • conducting fundamental data-driven science research through data collected from deployed AI systems, such as an AI-augmented learning platform to answer questions on human-AI collaboration, self-directed learning, intelligent knowledge discovery, and racial and social equity issues, and

  • co-optimizeing AI systems at the intersection of multiple levels, including the computing level with modern and emerging computing devices and paradigms, the algorithmic level with data, AI, machine learning and deep learning, and the application level centering around human and society.

My research style is a use-case driven, end-to-end approach to systematically tackle relevant research problems through a combined systems and AI innovations. I build functional AI solutions on existing AI systems, develop productivity tools to measure and quantify bottlenecks and issues in existing AI solutions, models and systems, and then co-optimize the end-to-end AI solutions through innovations in AI theories, algorithms, and computing systems.

Below are some projects I did that reflected the principles of such a systematic approach to conduct advanced cross-disciplinary research.

End-to-end AI Systems Research for Scalable Personalized Education

In the past five years, I became particularly interested in addressing one of the most important societal challenges, i.e., how to use AI to make personalized education affordable and accessible for all. In case you're wondering, the other equally important societal challenge I worked on is sustainability with renewable integration.

Through the collaboration with my former IBM colleagues,I understood that there were many research challenges that would benefit from a close collaboration between AI and systems.

Therefore, I have been pursing the following three-pronged research agenda together with my former IBM colleagues, such as those who developed Watson Education’s solutions for early education (with Sesame Street) and college tutoring (with Pearson Education), and my academic collaborators, such as Profs. Wen-mei Hwu, Deming Chen, Kevin Chang, Chengxiang Zhai, Yiyu Shi, Honghui Shi, Weiwen Jiang, Meng Wang etc.

  • The first one is a driving AI use case, titled Creative Experiential Learning Advisor (CELA), aims to push the boundaries of fundamental data science and AI research, and to address the complex interplay issues in designing large-scale AI solutions and hardware systems. CELA’s vision is to develop novel AI technologies that can assess students’ knowledge gaps, suggest personalized learning modules to students, analyze video streams to understand student behaviors and actions (for example, when they conduct the suggested science experiments in a particular setting), infer students’ emotions and state-of-learning, and eventually engage students proactively through conversations with a pedagogical AI agent. Some of my preliminary work on CELA included a technique to ground image cues to the corresponding textual cues (NIPS’18 Oral, ICCV’21), a way to map students’ curriculum concepts to different science projects (ACL’18), a video understanding technique to recognize students’ fine-grain actions (ICCV’19 Oral), and unsupervised NLP techniques to find high-quality phrases (SDM’21) and assess their semantic capacity (EMNLP’20, ACL’21).

  • The second research thrust, titled Cognitive Application Builder (CAB), aims to develop automated AI tooling and methodologies to identify AI applications’ across-stack computation, memory and networking bottlenecks on a distributed heterogeneous system. CAB tools will not only help co-optimize AI algorithms (e.g., model training, inference and deployment) and the underlying heterogenous hardware systems, but also boost AI development productivity for researchers who are not systems experts. Some research highlights on CAB included an across-stack profiling and benchmark platform to identify DNN models and DL frameworks’ performance bottlenecks on a heterogenous AI system (that won the Best Paper Awards at both IPDPS’20 and CLOUD’20, respectively), and an open-source tool for scalable characterizing communication and memory access patterns of an enterprise AI system (that won the ICPE’19 Best Paper).

  • The third research thrust, titled Near-Memory Acceleration (NMA) for AI hardware systems innovation, aims to design ground-breaking AI hardware systems and innovative methodologies to optimize existing hardware systems for big data AI solutions, where moving large volumes of data across heterogeneous computing nodes is slow and expensive. NMA tries to bring AI computation near (or in) memory with storage-scale capacity, to improve effective use of available bandwidth in a disaggregated and tiered computing environment, and to consider data persistence and programmability issues. Some sample work on NMA included a near-memory acceleration prototype based on an IBM proprietary memory extension card (that won the MICRO’18 Best Paper nomination), a near memory AI acceleration solution to support intelligent big-data queries (MICRO’19), and a hardware and software co-design approach for DNN design for resource-constrained embedded devices (MLSys’20, ICCAD’18 that won the Best Paper Award). Under the same thrust, I was recently became interested in understanding the impact of Quantum Computing on system accelerations (Nature Communications'21).

The three-proned research agenda together form a cohesive, forward looking AI systems research agenda that is still valid even for today.

Future of X-Lab: Accelerating AI Systems and Solutions Holistically

Going forward, I would like my research lab, X-Lab@UB, to continue to push such a holistic AI Systems research vision to the next level. I believe hybrid computing infrastructure that combines core clouds, edge clouds, embedded IoT, AR/VR devices, and even future Quantum Computers will be at the forefront of the next generation of AI Systems research, and the optimization of an AI system needs to be informed by application requirements and user preferences.

If you share such a research vision and are interested in joining me, please feel free to get in touch.