Projects

Multilevel neural simulation-based inference

Paper (NeurIPS 2025)

We suggest a method to combine multifidelity simulation in neural SBI. In many scientific domain, it is common to have simulators with different fidelity levels; low fidelity simulator simplifies reality a lot, or use coarse mesh size, while high fidelity simulator is more aligned with real world. Ideally, we would use many high-fidelity samples to do SBI; however it is often computationally prohibitive. In this paper, we suggest a method to combine these multifidelity simulator using the framework of multilevel Monte Carle (MLMC). We show that using MLMC to approximate the training objective of neural SBI lead cost efficient SBI.

@article{hikida2025,
  title={Multilevel neural simulation-based inference},
  author={Hikida, Yuga and Bharti, Ayush and Jeffrey, Niall and Briol, Fran{\c{c}}ois-Xavier},
  journal={Advances in Neural Information Processing Systems (NeurIPS)},
  year={2025},
  url={https://arxiv.org/abs/2506.06087}
  }

Amortised and provably-robust simulation-based inference

arXiv Preprint

We leverage the framework of generalised Bayesian inference (GenBayes) to achieve posterior estimation that is robust to outliers. Specifically, we propose Neural Score Matching Bayes, a method that uses the score matching divergence as the loss function within the GenBayes framework. In this approach, the parameters of an unnormalised likelihood are modelled by a neural network, enabling application in simulator-based settings.

@article{bharti2026amortised,
  title         = {Amortised and Provably-Robust Simulation-Based Inference},
  author        = {Bharti, Ayush and Dellaporta, Charita and Hikida, Yuga and Briol, Fran{\c{c}}ois-Xavier},
  year          = {2026},
  eprint        = {2602.11325},
  archivePrefix = {arXiv},
  primaryClass  = {stat.ML},
  doi           = {10.48550/arXiv.2602.11325}
}
Past projects

Lossless Visualization of 4D Compositional Data on a 2D Canvas

arXiv Preprint

The paper suggests the method to visualize 4D compositional data on a 2D canvas. This can be used to visualize data such as posterior model probabilities for 4 models on (2D) paper without loss of information and distortion. My contributions are general polishing of manuscript, creating 3D interactive figures and animation, and development of R package (planned).

R package “metabmc”

The package implements meta-uncertainty quantification for Bayesian model comparison. This can be used to evaluate trustworthiness of posterior model probabilities, which tend to take extreme values. Check the project website for more detail. The development is still in progress and will be published in near future.

R package for Bayesian proxy structural VAR

The package implements Bayesian vector autoregressive model with proxy variables for structural shocks in economy. This can help agents like government and central bank to make decisions, considering the expected effects in the system of economy. The development is still in progress and will be published in near future.