Invertible Memory Flow Networks

Liyu Zerihun, Alexandr Plashchinsky, [0.5em] VECTOR Labs[0.3em] [email protected] , [email protected] · 2026

Read Invertible Memory Flow Networks by Liyu Zerihun, Alexandr Plashchinsky, [0.5em] VECTOR Labs[0.3em] [email protected] , [email protected] (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-2602-00535.txt.

Description

Liyu Zerihun and Alexandr Plashchinsky tackle long sequence memory, where recurrent networks suffer vanishing gradients and Transformers scale quadratically. Their Invertible Memory Flow Networks make compression tractable by factorisation, splitting it into merges through a binary tree of sweeper modules. Each sweeper learns a simpler 2-to-1 task, giving logarithmic depth with sublinear error growth in sequence length. For online inference the team distils a recurrent student with constant sequential steps. Tests on MNIST sequences and UCF-101 videos confirm high-dimensional data can be compressed over long sequences.

Catalog record

CREATORLiyu Zerihun, Alexandr Plashchinsky, [0.5em] VECTOR Labs[0.3em] [email protected] , [email protected]
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE39815
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602.00535

Browse the full library · About the Tower of Babel