← Research site

Tianbai (Toby) Yu

yutianbai423@gmail.com · Champaign, Illinois · GitHub · Google Scholar · LinkedIn

Research interests

Learning objectives, representations, and architectures for reusable knowledge, multimodal and spatial reasoning, and generalization.

Education

University of Illinois Urbana-Champaign

B.S. in Statistics and Computer Science

Expected May 2027

GPA: 3.92 / 4.00

Technical skills

Fluent: Python, PyTorch · Familiar: TensorFlow, CUDA · Basic: C/C++

Publication

Shiji Zhou*, Tianbai Yu*, Zhi Zhang, Heng Chang, Xiao Zhou, Dong Wu, and Han Zhao. “Efficient Utility-Preserving Machine Unlearning with Implicit Gradient Surgery.” Advances in Neural Information Processing Systems 38, NeurIPS 2025, Main Conference Track. [Code] [Poster / video]

* Equal contribution.

Research experience

Efficient Utility-Preserving Machine Unlearning with Implicit Gradient Surgery

May 2024–October 2025

Research mentor: Shiji Zhou · Co-first author and NeurIPS 2025 poster presenter

  • Formulated competing objectives as constrained optimization and derived an implicit gradient-surgery update requiring one backpropagation per step.
  • Implemented and maintained the complete PyTorch codebase as the sole code contributor, including data, training, logging, and evaluation pipelines.
  • Led classification and image-generation experiments, ablations, analysis, and time-sensitive rebuttal experiments; co-wrote core parts of the reviewer response.

Self-Supervised Learning for 3D Scene Representation

June 2024–December 2025

Collaboration with Sipeng Zhang and members of NVIDIA Spatial Intelligence Lab / NuRec

  • Proposed masked multi-view feature prediction using DINOv3 features to learn cross-scene, view-consistent 3D scene representations.
  • Designed masking, camera-pose and spatiotemporal/positional embeddings, and a ViT-based predictor; implemented the end-to-end PyTorch pipeline.
  • Built a working prototype and integration interface and ran initial validation with collaborators.

Selected technical projects