Published November 6, 2025 | Version v2

ReLU Region Reason (Re3)

  • 1. ROR icon Mälardalen University

Description

ReLU Region Reason (Re3) has been introduced in this study. This method takes advantage of the piecewise-linear nature of ReLU networks to gain insights into how neurons are activated and accurately calculate how each feature contributes to the final
score and probability.

If you use this code in your research, please cite:
```bibtex
@article{Barua7328,
author = {Arnab Barua and Mobyen Uddin Ahmed and Shahina Begum},
title = {Mechanistic Interpretability of ReLU Neural Networks Through Piecewise-Affine Mapping},
pages = {1--35},
month = {January},
year = {2026},
journal = {Machine Learning},
url = {http://www.es.mdu.se/publications/7328-}
}
```
Barua, A., Ahmed, M.U. & Begum, S. Mechanistic Interpretability of ReLU Neural Networks Through Piecewise-Affine Mapping. Mach Learn 115, 17 (2026). https://doi.org/10.1007/s10994-025-06957-0
 

Files

accelerometer_gyro_mobile_phone_dataset.csv

Files (25.7 MB)

Name Size Download all
md5:dc8b9464ebcfd3a61ce5b4e1b0e7cb17
2.1 MB Preview Download
md5:8f90e44047182c5f32d10f0c4dd4890f
22.7 MB Preview Download
md5:341a3b5244f213282b7b0920b729c592
4.0 kB Preview Download
md5:fff2e355450d594c82c283855512ddd2
36.2 kB Preview Download
md5:4ffff04a198e9cdf2fb01f05ee0e7001
41.2 kB Preview Download
md5:f848be81662465406aec774298c56491
48.5 kB Preview Download
md5:698922d0acab3043687f29a2a9effba1
3.7 kB Preview Download
md5:f5b056f6834f00c5a8cc73c45694708c
140 Bytes Preview Download
md5:24f6f70fcf40fa115de5c71859dc7108
9.3 kB Preview Download
md5:e7368d48fa690b0331a09b430d55fb6a
702.9 kB Download
md5:0a71c3eccab2b35d9eaf5f4dee070bc5
6.4 kB Download
md5:4f4a688ed988164deb0b7c77a7135bfc
3.6 kB Download

Additional details

Funding

European Commission
FITDRIVE - Monitoring devices for overall FITness of Drivers 953432
European Commission
TRUSTY - TRUSTWORTHY INTELLIGENT SYSTEM FOR REMOTE DIGITAL TOWER 101114838

Software

Programming language
Python