Efficient Utility-Preserving Machine Unlearning with Implicit Gradient Surgery
An efficient constrained-optimization approach to machine unlearning that preserves utility while requiring one backpropagation per step.
* Equal contribution.
Statistics + Computer Science · UIUC ’27
I am an undergraduate at the University of Illinois Urbana-Champaign studying Statistics and Computer Science. I am broadly interested in machine learning and am still exploring the questions I want to pursue most deeply. So far, my research has focused on machine unlearning and self-supervised 3D scene representation, combining clear algorithmic ideas with careful experiments.
01
Publication
02
Research
Self-Supervised Learning for 3D Scene Representation
Collaboration with NVIDIA Spatial Intelligence Lab / NuRec
Proposed masked multi-view feature prediction to learn view-consistent 3D scene representations from posed images, and built the end-to-end PyTorch prototype.
03
Selected technical projects
Systems
GPT-2 & FlashAttention CUDA Kernels
Implemented GPT-2 inference and FlashAttention kernels in CUDA/C++, then validated outputs against a reference implementation.
Computer vision
ImageNet CNN Training Pipeline
Built an end-to-end PyTorch pipeline and trained an AlexNet-style CNN on 10% of ImageNet-1k using a single 6 GB laptop GPU.
View code ↗04
Background
Education
University of Illinois Urbana-Champaign
B.S. in Statistics and Computer Science · Expected May 2027
GPA: 3.92 / 4.00
Technical
Python · PyTorch
TensorFlow · CUDA · C/C++