Thanks 🙏

Going forward, I'm excited to continue applying probabilistic ML tools to science, and to advance the foundations of generative modeling and black-box inference.

I'll be at #NeurIPS2025 -- happy to connect and chat!

🔗 www.dianacai.com

🧩 Generative modeling under misspecification

I study how generative models behave when their assumptions fail, and develop frameworks that remain robust under model mismatch—from mixtures and networks to nonparametric models.

🔬 Data acquisition for expensive scientific workflows

I design methods that guide what to measure next in costly experiments and simulations, e.g., in materials design, including physics-aware active search algorithms to accelerate stability predictions, and in genomics.

🧠 Black-box inference

I develop black-box probabilistic inference algorithms, including:
* Fast, flexible variational inference via score matching
* MCMC including for costly scientific simulations
* Simulation-based inference in misspecified & hierarchical settings

We develop an efficient VI algorithm using score matching, which reduces inference to a *non-negative least squares problem* with linear constraints -- can be solved efficiently!

Link to paper: arxiv.org/abs/2510.21598

Joint work with Robert Gower, David Blei, and Lawrence Saul.

Enter the Feynman identity, originally developed for loop integrals in quantum field theory:

It expresses a product of multiple fractions as an integral over the simplex.

➡️ The PoE becomes a continuous mixture of t's & then gives us a way to estimate Z and sample from the PoE

We construct a variational family that's a weighted product of multivariate t-experts. It captures skew, heavy tails, and multi-modality.

Products of experts (PoEs) are powerful -- but the normalizing constant Z is usually intractable and sampling is hard.