Check out the paper below, or come chat to us in Vancouver for more details!

Paper: lnkd.in/eHjtXYbx

#AI4Science #Bayes #ICML #simulation

We propose and compare two ways to learn RPs for arbitrary simulation models using neural mutual information estimators, showing that these approaches can accurately recover known reference priors and provide a way to subsequently perform neural simulation-based Bayesian inference *for free*.

Finding reference priors is difficult, however – this is especially true for simulation models, which are used across scientific domains (particle physics simulators in physics, agent-based models in economics, etc.).

This is measured by a KL divergence between the posterior and prior. The idea is that these priors provide a means to conduct prior sensitivity analyses, allowing modellers to measure how much prior information is incorporated into their Bayesian procedures through the prior distribution they chose.

RPs were developed as one possible way to formalise the notion of “minimal prior information” in Bayesian inference, and are priors that maximise the expected information to be gained from data about the model’s parameters.

@neuripsconf.bsky.social

@inetoxford.bsky.social @inet-complexity.bsky.social @compscioxford.bsky.social #causal #simulation #Bayes

2) “Model Exploration through Marginal Likelihood Entropy Maximisation”, accepted at the #NeurIPS2024 D3S3 Workshop. We propose a method for exploring the behaviours a (differentiable) simulator can generate – useful for understanding your mechanistic model! openreview.net/pdf?id=dPszK... (3/3)

This provides decision- and policy-makers with a way to reduce the cost of running simulations when designing policies and interventions. Useful in time-critical settings, such as emerging pandemics! (2/3)