Brown CS News

Anh Truong And Qiuhong Anna Wei Win The Randy F. Pausch Computer Science Undergraduate Summer Research Award

Click the links that follow for more news items about Peter Norvig, the Randy F. Pausch '82 Computer Science Undergraduate Summer Research Award, and other recent accomplishments by our students.

The Randy F. Pausch '82 Computer Science Undergraduate Summer Research Award, given this year to Anh Truong and Qiuhong Anna Wei to support their work with Brown CS faculty members Daniel Ritchie and Srinath Sridhar, respectively, recognizes strong achievement from undergraduate researchers and offers them the opportunity to continue their work over the summer.

A generous gift from Peter Norvig '78 (a Director of Research at Google and a thought leader in the areas of artificial intelligence, natural language processing, information retrieval, and software engineering) established the award, which provides $10,000 annually to support an undergraduate engaged in an intensive faculty-student summer research partnership. The gift honors the life and work of Randy F. Pausch '82, a renowned expert in computer science, human-computer interaction, and design who died of complications from pancreatic cancer in 2008. "His story is inspiring," Peter says, "and this is an opportunity to remember him."

Anh Truong

Anh explains that his research lies at the intersection of computer graphics and machine learning and deals with finding ways to generate novel yet plausible 3D shapes by creatively borrowing geometric features from given shapes. "Our work," he says, "focuses on an emerging class of shape representations called neural implicits, where shapes' surfaces are encoded implicitly as the level sets of learned functions. Neural implicits are compelling, because unlike traditional ways of representing geometry such as meshes or voxel grids, they 'decouple' resolution from memory costs: all that's needed to perfectly describe the underlying geometry of a shape at a theoretically infinite resolution are the weights of a small neural network."

The ability of GPT and its derivatives to synthesize human-like text was an inspiration, Anh says, particularly their strong few-shot learning abilities, where a user can provide examples of desired output text and prompt the model to synthesize new text in a similar vein. His project aims to achieve an analog with 3D models rather than natural language. Specifically, he hopes to explore how autoregressive models can be used to achieve few-shot synthesis of implicit shapes.

"I first encountered Daniel," Anh tells us, "through his past CSCI 1470 lecture recordings. His passion for the subject was enough to hook me into joining the lab. After completing the lab's starter project, I was surprised to have so much freedom in my choice of research. I never would have expected a professor to entrust undergraduates to lead research projects. What I enjoy the most about working with Daniel is the completely open environment and sense of adventure he fosters in the lab. I didn't expect my understanding of failure to be so turned on its head that I now enjoy discussing failed experiments in our meetings. I'm equally as grateful to have had the chance to work with my incredible labmates: Paul Biberstein, Kenny Jones, and Kai Wang."

Anna Wei

"Canonicalization, or recognizing and resolving to a standard and ‘normal’ form," explains Anna, "is an important aspect of reasoning about, representing, and interacting with the physical world. For 3D scene understanding and manipulation, canonicalization of object arrangement is a key problem. Most prior research has focused on what is regular, through tasks such as scene synthesis and scene rearrangement. This produces compelling regular scene arrangements but lacks the ability to directly work with abstracted principles defining canonical scenes, which limits their usefulness and our control over them."

For this project, Anna and her collaborators aim to answer why a scene is regular by directly handling the principles governing canonical arrangements. This could offer multiple new advantages: low domain dependence and better generalizability, control during synthesis or rearrangement, and insights into new ways to represent scenes. There are two main challenges: (1) a lack of general representation for the various regularities that contribute to canonicalization, and (2) the difficulty of learning regularity rules in isolation due to the many-to-many relation between regularity principles and scenes. To address these, the project will have two stages: (1) defining representations of regularity rules and (2) combining the task of learning regularity rules with their existing regular rearrangement framework to develop a method that has both "what" and "why" capabilities.

"This project extends from our recent submission to CVPR," says Anna, "and I'm excited to see where we can go next! Research has been exceptionally rewarding but also challenging at times. As Einstein once put it, 'If we knew what we were doing it would not be called research.' As much as I'm attracted to the uncertainty, constant learning, and the important feeling of 'being stupid', it would have been extremely difficult without the support of others. I'm beyond grateful for the guidance and support of Srinath and my collaborators, and I look forward to working with them over the summer!"

Anh and Anna's excitement and curiosity are exactly what Peter Norvig is looking for. He sees this award as a multiplier that will amplify the value of his gift and extend it through time. "In the past," he says, "we had to build all our own tools, and we didn't have time to combine computer science with other fields. Now, there are so many opportunities to do so. I think it's a wise choice: you invest in things that you think will do good, and educating a student allows them to help add to the things that you're already trying to accomplish."

For more information, click the link that follows to contact Brown CS Communications Manager Jesse C. Polhemus.