rbt4dnn : Requirements-based Testing of Neural Networks

Nusrat Jahan Mozumder, Felipe Toledo, Swaroopa Dola, Matthew B. Dwyer · 2025

Read rbt4dnn : Requirements-based Testing of Neural Networks by Nusrat Jahan Mozumder, Felipe Toledo, Swaroopa Dola, Matthew B. Dwyer (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-2504-02737.txt.

Description

Mozumder, Toledo, Dola, and Dwyer present a way to test neural networks against written requirements, in this two thousand twenty-five paper. Ordinary software testing checks stated requirements, but neural networks learn functions nobody can fully write down, so the method breaks.

Their approach, RBT4DNN, turns natural language requirements into combinations of named semantic features. A generative model, fine-tuned on training examples matching those features, produces realistic test inputs that meet each requirement's conditions. Running the network on those inputs shows whether its outputs obey the stated postconditions. Tests came out realistic and diverse in evaluation. The method can hunt faults and, during development, show engineers where a model's behavior drifts from the spec.

Catalog record

CREATORNusrat Jahan Mozumder, Felipe Toledo, Swaroopa Dola, Matthew B. Dwyer
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE92 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2504-02737

Browse the full library · About the Tower of Babel