Research Scientist at Meta
New York, NY, USA
Joined October 2022
- Tweets32
- Following93
- Followers115
- Likes52
Pinned Tweet
Excited to share our ICML 2025 paper:
"Scalable Gaussian Processes with Latent Kronecker Structure"
We unlock efficient linear algebra for your kernel matrix which *almost* has Kronecker product structure.
Check out our paper here:
arxiv.org/abs/2506.06895
Jihao Andreas Lin retweeted
1/14 Is Muon “better” than Shampoo?
We argue that their relationship parallels Adam's relationship with Signum. Analogous to @lukas_balles and Hennig’s (2018) decomposition of Adam into element-wise scaled Signum, we can decompose Shampoo as left- and right-adapted Muon.
Jihao Andreas Lin retweeted
1/9 In practice, the Shampoo optimizer crucially relies on several heuristics.
In our NeurIPS 2025 spotlight paper, we investigate the role of learning rate grafting and infrequent preconditioner updates in Shampoo by decomposing its preconditioner.
arxiv.org/abs/2506.03595
We demonstrate that Latent Kronecker Gaussian Processes (LKGPs) outperform state-of-the-art sparse and variational GPs on real-world datasets with up to five million examples, including robotics, learning curve prediction, and climate modeling tasks.
Big thanks to my amazing collaborators: Sebastian Ament, @MaxBalandat , @davidmeriksson , @jmhernandez233 , and @eytan !
If you are attending ICML 2025, feel free to check out our poster and have a chat with us!
Thu 11:00am - 1:30pm in East Exhibition Hall A-B, # E-1308
Jihao Andreas Lin retweeted
Exited to share our new paper accepted by ICML 2025
👉 “PTSD: Progressive Tempering Sampler with Diffusion”
, which aims to make sampling from unnormalised densities more efficient than state-of-the-art methods like parallel tempering.
Check our threads below 👇
Jihao Andreas Lin retweeted
Looking for Postdoc/Research Assistant to work on deep generative models, with a focus on the domain of molecules. Application deadline 3 July 2025. More info: jobs.cam.ac.uk/job/51666/
Jihao Andreas Lin retweeted
New revision of “BNEM: A Boltzmann Sampler Based on Bootstrapped Noised Energy Matching” 🚀
(1/6) We introduce NEM and BNEM — diffusion&energy-based neural samplers for Boltzmann distributions.
📝 Read our paper at [arxiv.org/pdf/2409.09787]
Jihao Andreas Lin retweeted
1/7 Still using Adam?
If anyone wants to try a distributed PyTorch implementation of SOAP/eigenvalue-corrected Shampoo with support for low precision data types instead, here you go.
github.com/facebookresearch/…
Jihao Andreas Lin retweeted
Diffusion models are so ubiquitous, but it's difficult to find an introduction that is concise, simple and comprehensive.
My supervisor Rich Turner (with me & some other students) has written an introduction to diffusion models that fills this gap:
arxiv.org/abs/2402.04384
Jihao Andreas Lin retweeted
How can we get VLMs to help robots solve complex long-horizon tasks? Introducing VisualPredicator: an agent that leverages VLMs to learn predicates and operators for classical planners. Our system can stack blocks, balance weights on a balance beam, and even pour coffee🦾!
[1/9]
"Improving Linear System Solvers for Hyperparameter Optimisation in Iterative Gaussian Processes"
Three techniques to accelerate marginal likelihood training in GPs by up to 72x without sacrificing performance!
Check out our paper here: arxiv.org/abs/2405.18457
(1/6)
With large datasets, solvers require too much compute to converge. Instead, they are stopped after a limited compute budget.
Combining early stopping and warm starting, solvers accumulate progress across marginal likelihood steps, amortising the costs of linear solves!
(5/6)
Many thanks to my collaborators @shreyaspadhy, @kayembruno, @JaviAC7, and @jmhernandez233 for this amazing project!
Check out our paper here: arxiv.org/abs/2405.18457
(6/6)