News
2026
- Collaborated with Scale AI on ResearchRubrics (ICLR 2026), PRBench (ACL Main 2026), and the Remote Labor Index. Check them out!
2025
- Selected as a 2025 Stanford Rising Star in Data Science, a workshop co-organized by Stanford Data Science, Halıcıoğlu Data Science Institute (UC San Diego), and the University of Chicago. Presented on the interplay between efficiency and privacy in distributed learning. See this year's cohort here.
- Sat on the CFAgentic@ICML'25 panel with Niloofar Mireshghallah, William Lindskog-Münzing, and Chi Wang to discuss: "Will the future of AI agents be collaborative and federated?"
- Presented "Mitigating Unintended Memorization with LoRA in Federated Learning for LLMs" (joint with EPFL) as an oral at both the MemFM and CFAgentic workshops @ ICML 2025.
- July: After an exciting year at Yale working on applied ML, I am now at the University of Chicago as a Postdoctoral Scholar, co-hosted by Tian Li and Ce Zhang.
- January: Rotated to Dr. Hoon Cho's lab at Yale. Continuing work in low-resource, private, distributed ML for proteomics and drug discovery.
- "HashEvict: A Pre-Attention KV Cache Eviction Strategy using LSH" is now on arXiv. Pushing the Pareto frontier on speed and output quality for KV compression.
2024
- "Llama-3-Meditron: An Open-Source Suite of Medical LLMs" accepted as an oral at the AAAI'25 Workshop AI2ASE.
- Co-hosted our watermarking workshop at NeurIPS'24. Catch the event here!
- Presented distributed learning platform "DISCO" at GDHF'24 in Nairobi, Kenya (joint with EPFL).
- Two works accepted to the NeurIPS'24 Workshop on ML and Compression: "Shrinking the Size of Deep Extreme Multi-Label Classification" and "LSH-E Tells You What To Discard."
- Competition proposal "Erasing the Invisible: A Stress-Test Challenge for Image Watermarks" accepted by NeurIPS'24.
- "WAVES: Benchmarking the Robustness of Image Watermarks" accepted to ICML'24.
- Headed to Yale LiGHT for a postdoc on LLM compression and FL for healthcare analytics.
- Selected as a 2024 RSA Security Scholar for work in privacy-preserving ML.
2023
- Paper accepted to NeurIPS'23: "Large-Scale Distributed Learning via Private On-Device Locality-Sensitive Hashing." Also presented at the ICLR 2023 Workshop on Sparsity in Neural Networks.
- Presented at SLowDNN'23 in Dubai: "A Framework for Representation and Fast Evaluation of Multilinear Operations in Convolutional TNNs."
2022
- Paper accepted to ICLR'23: "SWIFT: Rapid Decentralized Federated Learning via Wait-Free Model Communication." Also received an abstract selection for the 2023 Qualcomm Innovation Fellowship and a spotlight at FL-NeurIPS'22.
- Paper accepted to NeurIPS'22: "SketchGNN: Scalable Graph Neural Networks," plus two papers at GL Frontiers 2022.
- Oral presentation for "Comfetch: Federated Learning of Large Networks on Memory-Constrained Clients via Sketching" at FL-AAAI-22.
- Paper accepted to MSML'22: "Practical and Fast Momentum-Based Power Methods."
2021
- Nominated as a 2022 Apple Scholar in AI/ML by the University of Maryland, College Park.
- Awarded an NSF COMBINE Fellowship for 2021–2022.
Selected Works
Mitigating Unintended Memorization with LoRA in Federated Learning for LLMs
We show that LoRA reduces memorization in federated LLM training by up to 10x without major performance loss, across high-risk domains (medicine, law, finance) and model sizes from 1B to 70B. LoRA composes naturally with other privacy techniques like gradient clipping, secure aggregation, and Goldfish loss for further protection.
A Technical Report on "Erasing the Invisible": The 2024 NeurIPS Competition on Stress Testing Image Watermarks
Key findings of our 2024 NeurIPS competition. Spoiler: the winners were really good at removing watermarks! Bonus: we've released all submissions to help further research on robust watermark design.
Benchmarking the Robustness of Image Watermarks
We systematically reveal weaknesses in modern image-based watermarking protocols, including generative approaches. Check out the benchmark at wavesbench.github.io.
Large-Scale Distributed Learning via Private On-Device LSH
Using a new family of hash functions, we develop one of the first private, personalized, and memory-efficient on-device LSH frameworks for training recommender DNNs on extreme multi-label datasets.
SWIFT: Rapid Decentralized Federated Learning via Wait-Free Model Communication
We enable wait-free model training for peer-to-peer FL using model caching. Provable convergence at SotA rate; empirically significantly speeds up global model convergence.
Sketch-GNN: Scalable Graph Neural Networks with Sublinear Training Complexity
A sketch-based algorithm whose training time and memory grow sublinearly w.r.t. graph size by training GNNs atop compact sketches of graph adjacency and node embeddings.
Constructions of Difference Sets in Nonabelian 2-Groups
We determine that all groups of order 256 not excluded by the two classical nonexistence criteria contain a difference set, resolving a 25-year-old question posed by John Dillon.
Practical and Fast Momentum-Based Power Methods
We provide novel momentum-based power methods, DMPower and DMStream. In contrast with prior art, these accelerated methods do not depend on spectral knowledge.
Fast GPU Convolution for CP-Decomposed Tensorial Neural Networks
A GPU algorithm for convolution with decomposed tensor products. Up to 4.85x faster execution than cuDNN for some tensors.
Nonabelian Orthogonal Building Sets
We examine recent construction techniques of Hadamard difference sets in 2-groups and an extension of orthogonal building sets to nonabelian groups.
Unique Minimal Forcing Sets and Forced Representation of Integers by Quadratic Forms
We use Bhargava's theory of escalators to establish infinite families of positive integers without unique minimal forcing sets.
Improving the Error-Correcting Code Used in 3-G Communication
We give a construction of a [30,10,11] non-cyclic code that improves upon the [30,10,10] code described in Samsung's patent US 7706348.
Artwork
I enjoy creating photorealistic artwork in my free time.
