RT @nvimorozov: (1/n) The usual assumption in GFlowNet environments is acyclicity. Have you ever wondered if it can be relaxed? Does the ex….

RT @QueerinAI: 1/ 💻 Queer in AI is hosting a social at #ICML2025 in Vancouver on 📅 July 16, and you’re invited! Let’s network, enjoy food a….

Relative clause attachment ambiguity not intentional ;).

An oasis of inclusive science and solidarity amid the monotonically increasing NeurIPS madness that I'm proud to be supporting in a small role this year.

A great pleasure to crash two Bayesian statistics conferences with a dose of diffusion wisdom — last week in Singapore (, now in Cambridge ( — with the two authors of this very nice paper.

@siddarthv66 @mh_steps Paper here:

Great paper by @siddarthv66, @mh_steps, et al. on amortised inference in latent spaces of generative models, generalising our past work (. Useful for alignment, planning in latent space, inference in probabilistic programs?.

RT @josephdviviano: Ecstatic to show off some work my brilliant colleagues and I did at @iclr_conf this year! 🚀. We address the credit assi….

RT @sarthmit: 🚀 New Preprint! 🚀. In-Context Parametric Inference: Point or Distribution Estimators?. Thrilled to share our work on inferrin….

RT @TristanDeleu: My PhD thesis entitled "Generative Flow Networks: Theory and Applications to Structure Learning" is now available on Arxi….

RT @MarcinSendera: Happy to share one of my last works! If you are interested in diffusion samplers, please take a look🙃! Many thanks for a….

This delightful collaboration built upon my past work with @MarcinSendera @jarridrb ( and that of the brilliant @julberner and @lorenz_richter ( . Thanks to all! 9/9.

Have a look at our code here. Multiple objectives, exploration strategies, and time discretisation strategies are implemented in a common framework. 8/9.

We are eager to see extensions of this work to non-Markovian sequential generation, discrete state spaces, and posterior sampling under diffusion priors, as well as discretisation error and generalisation bounds. Numerical analysis and stochastic calculus are key tools here! 7/9.

The fact that these objectives are well-behaved asymptotically justifies the use of coarser (perhaps non-uniform) time discretisations during training than during sampling. This leads to greatly improved sample efficiency and even allows the use of time-local objectives. 6/9

But all of these objectives run in a time discretisation. What happens when we take the step size to zero? For each objective, we prove that the limit is a well-understood continuous-time object. For example, the detailed balance condition approaches the Fokker-Planck PDE. 5/9

Such time-reversal can be enforced using a zoo of objectives computed through differentiable simulation (connected to stochastic control) or off-policy divergences (connected to entropy-regularised RL). We show connections among these objectives. 4/9

Diffusion models are usually trained to maximise a variational bound on data likelihood, but they can also be fit without any data, by enforcing time-reversal of the generative and noising processes, where the latter has a target density as its boundary condition. 3/9

Inference, or sampling, in high-dimensional spaces is important in Bayesian statistics and in scientific applications (inverse problems, molecular dynamics, . ). Problem: How can we train diffusion models to amortise the cost of sampling a black-box target density? 2/9.

Happy to share our latest work on #diffusion models without data: building theoretical bridges between existing methods, analysing their continuous-time asymptotics, and showing some cool practical implications. #MachineLearning 1/9