Kevin Lam

Stats/ML PhD student @OxfordStats

[Tyler Farghly @tylerfarghly]

I'm at @ISBA_events in Nagoya this week. Come say hi and check out our work!

🔶Metropolis Adjusted Diffusion Models
Wed (poster w/ @new2_yj @khflam)
🔶Rapid mixing of stereographic MCMC for heavy-tailed sampling
Fri & Sun (talk by Federica Milinanni w/ @new2_yj)

[Oxford Young Statisticians Seminar @OxfordYss]

This week, the @OxfordYss is attending the 2026 World Meeting of the @ISBA_events. An impressive number of our members are in Japan presenting their work.

If you are in Japan this week and would like to discuss their research, please reach out and connect! 🤝

[MolSS Reading Group @MolSS_Group]

How to incorporate Metropolis-Hastings into score-based Diffusion models to improve generation?

On the coming Tuesday, June 16th 4pm-5pm UK time, we will have @khflam to talk about "Metropolis-Adjusted Diffusion Models" (https://t.co/DHmrKlXWQM) 🚀

Join us with Zoom!! Links 👇

[Luke Heeney @heeney_luke]

We've lost an absolute giant today. RIP Dimitri Bertsekas. His probability and optimization books got me through my masters. Massive loss for the MIT community and the field.

[Tyler Farghly @tylerfarghly]

Check out the preprint:
https://t.co/qiucn15OX8

Huge thanks to my brilliant co-first authors @kevinhflam & @Chr1sW1lliams, as well as the fantastic @new2_yj, @yeewhye & @ArnaudDoucet1
[3/3]

[Tyler Farghly @tylerfarghly]

We propose two approaches:
🔵Exact Barker corrector via a two-coin algorithm: the first exactly adjusted corrector for score-based models
🟢Simpson's rule approximation: efficient with order 5/2 accuracy

Consistent FID gains on FFHQ, AFHQv2 & ImageNet-64 without retraining
[2/3]

[Tyler Farghly @tylerfarghly]

[📄preprint] Diffusion models 🤝 MCMC !

Diffusion model samplers are biased due to discretisation

💡The fix: Metropolis-type adjustment on corrector steps
❗️Challenge: no access to the density ratio, only the score
🔑Insight: the score (and some maths) is all you need...
[1/3]

[Oxford Young Statisticians Seminar @OxfordYss]

That’s a wrap on the #NeurIPS2025 Main Track! 🌴🏁

We presented 8 papers & loved the energy in San Diego. Huge thanks for the discussions! ☀️

But we aren't done. 🚀 Catch 4 more papers at the Workshops this weekend.

#NeurIPS @OxfordStats

[Thang Bui @thdbui]

At #NeurIPS2025 presenting 2 papers:

Rao-Blackwellised Reparameterisation Gradients with a rising star @kevinhflam, George Deligiannidis, and @yeewhye

Sparse Gaussian Processes: Structured Approximations and Power-EP Revisited with the inimitable Michalis Titsias

[Kevin Lam @khflam]

Paper: https://t.co/4v978RHYly

[Kevin Lam @khflam]

Excited to share our #NeurIPS2025 paper: Rao-Blackwellised Reparameterisation Gradients!

We propose R2-G2 as a general-purpose gradient estimator for latent Gaussians and as the Rao-Blackwellisation of reparam gradients.

Joint work with @thdbui @GeorgeDeligian9 @yeewhye

[Kevin Lam @khflam]

This explains why local reparam gradients have lower variance, but the local reparam trick is only applicable to independent pre-activations.

To lower gradient variance in more general cases, we use conjugate gradients to invert covariance matrices in backprop: yielding R2-G2!

[Leo Zhang @leoeleoleo1]

Wrote up some notes providing an introduction to discrete diffusion models, going into the theory of time-inhomogeneous CTMCs via generators/time-evolution systems.

What motivated me was the sheer difficulty of finding a useful reference which laid out the theory (e.g. Kolmogorov equations etc.) in a single place. Hope these can be helpful to others.