I will be at my first in-person NeurIPS, presenting 3 posters at the main conference (🧵)! Please get in touch to chat about - . - Diffusion models, sampling and conditional generation.- BayesOPT, GPs and BNNs. P.S. I'll be on the job market early next year, please reach out!.

Checkout this paper with some really interesting insights led by the excellent @JiajunHe614 and @YuanqiD - . TLDR: Neural density samplers really need guidance imposed through Langevin annealing to make them work well.

Thanks for the kind words @ArnaudDoucet1 ! I wanted to shout-out some other great work in the same vein as us - . (@julberner, @lorenz_richter, @MarcinSendera et al).(@msalbergo et al).(@junhua_c et al).

RT @ArnaudDoucet1: Tweeting again about sampling, my favourite 2024 Monte Carlo paper is by F. Vargas, @shreyaspadh….

Come chat with folks from our group!.

RT @CambridgeMLG: We're excited to be at #NeurIPS Vancouver!. See the papers we'll be presenting at the main conference below: https://t.co….

RT @AtinaryTech: Atinary @ #NeurIPS in Vancouver this week🍁 .Connect with our #AI #ML researchers @VictorSabanza & @shreyaspadhy. Our rese….

Finally, we're presenting a new symmetry-aware generative model that discovers which (approximate) symmetries exist in data for improved data efficiency. Catch co-author @JamesAllingham on Friday, Dec 13, 06:00 at Poster Session 6 East, #2500.

Second, we're presenting our work on speeding up marginal likelihood estimation in GPs by up to 72x without sacrificing performance! Catch co-author @JihaoAndreasLin on Thursday, Dec 12, 16:30 at Poster Session 4 East, #3910.

First, we're presenting our paper on efficient fine-tuning of diffusion models for conditional generation. Find us (w/ @DenkerAlexander, Francisco Vargas) on Thursday, Dec 12, 11:00 at Poster Session 3 East, #2500.

RT @JamesAllingham: I'll be at NeurIPS next week, presenting our work "A Generative Model of Symmetry Transformations." In it, we propose a….

We’re looking forward to chatting with folks at NeurIPS this year about this work! Work done with excellent collaborators @DenkerAlexander, Francisco Vargas, @DidiKieran, @SimMat20, @vdutor, @BarbanoRiccardo, @MathieuEmile, @julia_tweeting_, and @pl219_Cambridge!.

In conclusion, DEFT is a fine-tuning framework which.(1) leverages pre-trained models, reducing training time and computational resources, and .(2) is applicable to various conditional generation tasks, both linear and non-linear. (11/11).

Further, we apply DEFT to the protein design task of motif scaffolding. Here, we obtain good results with a h-transform model only 4% to 9% of the size of the original model, drastically reducing the training time. (10/11)

We show promising results both on linear imaging tasks (inpainting, deblurring, super-resolution and computed tomography) as well as non-linear imaging, including non-linear deblurring and phase retrieval. (9/11)

Naively conditioning SDEs can lead to a bias due to choosing the incorrect initial distribution. We show how to address this for the VP-SDE and connect it to recent work (Domingo-Enrich et al., 2024) that links this bias to the value function from stochastic control. (8/11).

Further, the connection to prior work on stochastic optimal control motivates us to use similar network architectures and encode the likelihood term as an inductive bias. This parameterization allows for a fast training of the model! (7/11)

We demonstrate that DEFT can also be interpreted as a stochastic optimal control problem, enabling the use of variational inference techniques for learning the h-transform with only a single noisy observation. (6/11).

The Offline Score-Matching loss requires a small dataset for fine-tuning. For Image Inpainting on ImageNet we obtain SOTA results with only 200 additional training images. (5/11)

In particular, we propose two loss functions: Offline Score-Matching and Online Stochastic Control. Both loss functions only require access to the forward pass of the pre-trained model and are agnostic to the underlying network architecture. (4/11)