Community Connections Events, Parents and Families Events Page 1

Unlike most machine learning applications, robotics involves physical constraints that make off-the-shelf learning challenging. Difficulties in large-scale data collection and training present a major roadblock to applying today’s data-intensive algorithms. Robot learning has an additional roadblock in evaluation: every physical space is different, making results across labs inconsistent.

Two common assumptions of the robot learning paradigm limit data efficiency. First, an agent typically assumes isolated environments and no prior experience. Second, agents frequently rely on vision alone to learn tasks. Victoria Dean will present work that lifts these two assumptions, improving the data efficiency of robot learning by leveraging multimodality (images and audio) as well as pretraining. She will then introduce a real-world manipulation benchmark for evaluating the generalization of policy learning methods via shared data and hardware.

Finally, Dean will turn to another robot learning application at a larger scale: optimizing heating and cooling systems in our buildings. Buildings account for nearly 40% of carbon emissions in the U.S., yet most building control policies fail to generalize beyond a single building. She will discuss ongoing work to benchmark generalization via the Smart Buildings Challenge, highlighting how machine learning can tackle pressing sustainability challenges.

Speaker

Victoria Dean is a teaching professor in the Department of Computer Science at Princeton University. Her research interests are at the intersection of machine learning and robotics, investigating generalization, benchmarking and democratization. She teaches across introductory computer science, machine learning and ethics and enjoys implementing real-world projects in my courses. Dean is passionate about teaching and mentoring undergraduate students, with an emphasis on improving diversity, equity, and inclusion.

Prior to joining Princeton, Dean was an assistant professor of computer science at Olin College of Engineering. She did her PhD at CMU’s Robotics Institute, advised by Abhinav Gupta and supported by the NSF GRFP and Siebel Scholars. She got her B.S. in computer science from MIT and spent time at DeepMind, Waymo, Deep Genomics, Counsyl, Google and Coursera.

CS Research Talk– “Rooting Generative AI in Cultural Knowledge: Artist-Centric Pipelines for Representing Diverse Hair Types on Screen,” Sarah Jobalia

Machine Learning often relies on massive amounts of data to capture relevant patterns. In contexts where experts are performing highly-skilled tasks, however, large datasets can be hard to create, and seemingly small miscalculations on the model’s part can be disproportionately problematic.

Sarah Jobalia, a Stanford University PhD candidate, will focus on the design of virtual 3D hair as a case study to show how grounding AI pipelines in cultural, scientific and social knowledge produces more realistic, diverse and reliable results. Virtual hair, which is both computationally complex and culturally important to represent faithfully, is central to adding depth and nuance to the stories we see on screen. In her research, she draws upon interdisciplinary insights to define strand texture as the various distinctive patterns that are formed by forces internal to a hair strand, presenting a novel taxonomy for rigorously and concisely describing texture. Mirroring social understandings of hair, Jobalia disentangles this texture from the hair’s overall style, which describes where the strands fall in relation to the head. Through learning and generating texture and style separately, she highlights two fundamental but separate aspects of hair, enabling generation that can capture the essence of desired hairstyles. This both yields fast generation of the curly and afro-textured hair types that previous works struggle to capture and allows for intuitive user interaction and compatibility with existing artist workflows. Broadly, these insights show how grounding AI tools in existing knowledge not only improves results but allows tools to support, rather than replace, experts.

Speaker

Sarah Jobalia is a PhD candidate in Stanford University’s Computer Science Department, advised by Ron Fedkiw. Her research centers on building more accessible tools for animation artists to represent characters of all backgrounds on screen. Specifically, through the study of hair- both on- and off-screen, Jobalia uses a combination of computer graphics and machine learning to automatically generate 3D hair styles from just a few images of any subject. By focusing on how hair type affects the way hair looks and moves, she builds tools that are equitably representative of all styles.

Submit an Event

Use the Submit Events form to add Mudd community events to the calendar.