GEM-FI: Gated Evidential Mixtures with Fisher Modulation
Marco Mustafa Mohammed, Fatemeh Daneshfar, Pietro Liò · 2026
Read GEM-FI: Gated Evidential Mixtures with Fisher Modulation by Marco Mustafa Mohammed, Fatemeh Daneshfar, Pietro Liò (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-2605-03750.txt.
Description
Marco Mustafa Mohammed, Fatemeh Daneshfar, and Pietro Liò introduce Gated Evidential Mixtures, a family of models that estimates uncertainty in a single forward pass. Evidential deep learning predicts Dirichlet evidence to express uncertainty, but it can stay overconfident and miss multimodal uncertainty about what the model does not know. Their design adds an internal energy signal that gates the evidence.
GEM-CORE learns a feature-level energy and maps it to a bounded gate that suppresses evidence when support is low. GEM-MIX adds a mixture of evidential heads with learned routing, and GEM-FI stabilizes the mixture with a Fisher-informed regularizer. Across image classification and out-of-distribution detection, the models improved calibration and separation; on CIFAR-10 the full version raised accuracy and sharply lowered Brier score.
Catalog record
| CREATOR | Marco Mustafa Mohammed, Fatemeh Daneshfar, Pietro Liò |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 96 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.03750