The Alpha-Alternator: Dynamic Adaptation To Varying Noise Levels In Sequences Using The Vendi Score For Improved Robustness and Performance

Mohammad R. Rezaei, Adji Bousso Dieng · 2025

Read The Alpha-Alternator: Dynamic Adaptation To Varying Noise Levels In Sequences Using The Vendi Score For Improved Robustness and Performance by Mohammad R. Rezaei, Adji Bousso Dieng (2025) 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-2502-04593.txt.

Description

Most models for data that unfold in time treat every step as if it were just as noisy as the rest. Mohammad Rezaei and Adji Bousso Dieng challenge that habit in this 2025 paper with a model they call the Alpha-Alternator. The model uses a diversity score to judge, at each step, whether the new input looks like noise or like real news.

A single learned dial sets the response. If new input that raises diversity counts as noise, the model leans on its memory of the past. If it counts as news, the model leans on the input instead. Training hides parts of the input to mimic noise. Across brain decoding and forecasting tests, the model beat both its parent design and strong state-space rivals.

Catalog record

CREATORMohammad R. Rezaei, Adji Bousso Dieng
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE39 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2502-04593

Browse the full library · About the Tower of Babel