The software implementation elegantly uses BayesFlow‘s modular data pipeline: - Observables are embedded by a summary network. - Context information (eg, prior and likelihood type) bypasses the summary net and enters the normalizing flow as direct conditions. 📀 Code: github.com/bayesflow-or...

The paper was led by @elseml.bsky.social, with multiple high-impact applications: 🦠 Disease outbreak modeling 🌎 Global warming thresholds 🧠 Human decision-making ✨ Sensitivity-aware amortized inference increases the amortization scope by a lot. Another step towards a Bayesian foundation model!

Hi, thanks for reaching out! In the context of amortized inference, it’s been shown that many of the algorithms we use are susceptible to adversarial attacks, and this can be mitigated by regularizing wrt Fisher information. 📝 Paper by @mackelab.bsky.social: arxiv.org/abs/2305.14984