Learning Encoding-Decoding Direction Pairs to Unveil Concepts of Influence in Deep Vision Networks

Alexandros Doumanoglou [email protected], Kurt Driessens [email protected], Dimitrios Zarpalas [email protected] · 2025

Read Learning Encoding-Decoding Direction Pairs to Unveil Concepts of Influence in Deep Vision Networks by Alexandros Doumanoglou [email protected], Kurt Driessens [email protected], Dimitrios Zarpalas [email protected] (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-2509-23926.txt.

Description

Deep vision nets seem to store ideas as rays in their inner space, but the code that writes and reads those rays is hidden. Alexandros Doumanoglou and 2 co-authors set out to find that code with no labels to guide them. For each idea, they posit two rays. One ray writes the idea into the code of an input. The other ray reads it back out.

Read rays come from grouping net outputs by their ray. Write rays come from signal vectors, read in a chance-based way. A trick that uses the net's own weights finds rays that sway the net's calls. On fake data with known truth, the method finds the true ray pairs. On real data, the read rays map to clean, single ideas.

Catalog record

CREATORAlexandros Doumanoglou [email protected], Kurt Driessens [email protected], Dimitrios Zarpalas [email protected]
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE216 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2509-23926

Browse the full library · About the Tower of Babel