Two UC San Diego Computer Science and Engineering (CSE) faculty members – Assistant Professor Jingbo Shang and Associate Professor Hao Su – were recognized with the National Science Foundation’s Faculty Early Career Development (CAREER) award. They join nearly 50 CSE recipients, a list that expanded earlier this year to include assistant professors Albert Chern, Tzu-Mao Li and Aaron Schulman.
The CAREER program offers NSF’s most prestigious award to support early-career faculty who have the potential to serve as academic role models in research and education and to lead advances in the mission of their department or organization. The program provides grants to support solutions-oriented research in science and engineering.
Assistant Professor Jingbo Shang, who is jointly appointed by CSE and Halıcıoğlu Data Science Institute (HDSI), received an NSF grant to continue his research in natural language processing (NLP) and, specifically, developing novel methods for text mining. Associate Professor Hao Su was awarded funding to research Embodied AI (Artificial Intelligence). Su is affiliated with HDSI as well as The Institute for Learning-enabled Optimization at Scale, the Artificial Intelligence Group, Contextual Robotics Institute, and the Center for Visual Computing.
Proving the Power of Extremely Weak Supervision
Shang’s CAREER project, Knowledge Extraction and Discovery from Massive Text Corpora via Extremely Weak Supervision, could address one of NLP’s fundamental challenges: how to extract and discover useful knowledge from massive text corpora with minimal user effort.
Shang proposes a new paradigm, extremely weak supervision (EWS), that builds on advances in neural language models (NLMs) to automate fundamental tasks in text mining, such as text classification, phrase mining, named entity recognition, and relation extraction. This method requires minimal input to define the task.
Shang hopes to unleash the full power of NLMs for EWS in both effectiveness and efficiency. He anticipates his research will yield algorithms and software applicable to a spectrum of data-intensive tasks, which is promising whether you are making decisions for the government, preparing a literature summary for scientific research, or running a small business.
“By using brief natural-language input instead of labor-intensive annotated training samples, this new paradigm will help democratize knowledge extraction and discovery, and extend its application beyond rich companies to ordinary, relatively untrained users with a broad range of needs,” said Shang.
Training AI to Adapt in the Real World
Adding to the rapidly evolving field of Embodied AI, Hao Su’s research project, Interaction-oriented 3D Representation Learning on Point Clouds, aims to train machines to see, reason, and interact with diverse environments. Ideally, these abilities would replicate the robust and adaptable visuomotor skills of humans.
For humans, visuomotor skills are acquired by practicing developmentally appropriate tasks; people learn by doing. Similarly, Su intends to leverage 3D point cloud data to build robotic systems that can learn from interaction experiences.
“My proposal plans to achieve the goal by building a large-scale simulation environment, training the visuomotor system in it, and applying sim2real transfer technology to deploy the robot in the real world,” said Su.
While Su anticipates transformative impacts across many domains – from manufacturing and civil engineering to agriculture, healthcare, and space exploration – his project focuses on four growing applications: smart assembly robots, AR devices to assist elderly and job training, exploratory robotics, and home robots.