Projects

Research Projects

Ensemble transform methods for neural network uncertainty

We propose ensemble transform methods that improve predictive uncertainty in neural networks. These methods are affine-invariant and applicable to any pre-trained network. The work uses ensemble transform Kalman filter-based MCMC methods for efficient posterior sampling and supports robustness against model misspecification.

📄 Read preprint

Bayesian inference of human language comprehension

Applied ensemble Kalman methods to simulate and infer uncertainty in the Sentence Gestalt model, a neural network model of human language comprehension. We used a scoring-rule-based Bayesian formulation to handle model misfit and explored inference robustness in probabilistic psycholinguistic models.

📄 Read preprint

Likelihood-free Bayesian inference in generative models

Developed novel sampling algorithms combining kernel-based scoring rules with ensemble transform MCMC samplers for high-dimensional generative models. Applications include dynamical systems, such as the stochastic Lorenz-96 model.

Ongoing & Miscellaneous

• to be added