Decoupled PFNs: Identifiable Epistemic–Aleatoric Decomposition via Structured Synthetic Priors
Richard Bergna, Stefan Depeweg, Jose Miguel Hernández-Lobato · 2026
Read Decoupled PFNs: Identifiable Epistemic–Aleatoric Decomposition via Structured Synthetic Priors by Richard Bergna, Stefan Depeweg, Jose Miguel Hernández-Lobato (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-06413.txt.
Description
Richard Bergna, Stefan Depeweg, and José Miguel Hernández-Lobato work on separating two kinds of uncertainty in Prior-Fitted Networks. These networks learn Bayesian prediction from synthetic tasks, but their output mixes epistemic uncertainty, which more data can reduce, with aleatoric noise, which it cannot. Decisions like active learning need the first kind only.
The authors first show the split cannot be recovered from predictions alone. Because they control the synthetic generator, they can label the true signal and noise for each query and train separate heads for each. Using only epistemic uncertainty for acquisition avoids wasted exploration in noisy settings and improves results.
Catalog record
| CREATOR | Richard Bergna, Stefan Depeweg, Jose Miguel Hernández-Lobato |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 98 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.06413