Forward Composition Propagation for Explainable Neural Reasoning

Isel Grau; Gonzalo Nápoles; Marilyn Bello-García; Yamisleydi Salgueiro; Agnieszka Jastrzębska

doi:10.1109/MCI.2023.3327834

Forward Composition Propagation for Explainable Neural Reasoning

Isel Grau (Corresponding author)
, Gonzalo Nápoles
, Marilyn Bello-García
, Yamisleydi Salgueiro
, Agnieszka Jastrzębska

Research output: Contribution to journal › Article › Academic › peer-review

6 Citations (Scopus)

217 Downloads (Pure)

Abstract

This paper proposes an algorithm called Forward Composition Propagation (FCP) to explain the predictions of feed-forward neural networks operating on structured classification problems. In the proposed FCP algorithm, each neuron is described by a composition vector indicating the role of each problem feature in that neuron. Composition vectors are initialized using a given input instance and subsequently propagated through the whole network until reaching the output layer. The sign of each composition value indicates whether the corresponding feature excites or inhibits the neuron, while the absolute value quantifies its impact. The FCP algorithm is executed on a post-hoc basis, i.e., once the learning process is completed. Aiming to illustrate the FCP algorithm, this paper develops a case study concerning bias detection in a fairness problem in which the ground truth is known. The simulation results show that the composition values closely align with the expected behavior of protected features. The source code and supplementary material for this paper are available at https://github.com/igraugar/fcp.

Original language	English
Article number	10384511
Pages (from-to)	26-35
Number of pages	10
Journal	IEEE Computational Intelligence Magazine
Volume	19
Issue number	1
DOIs	https://doi.org/10.1109/MCI.2023.3327834
Publication status	Published - Feb 2024

Funding

The work of Marilyn Bello was supported by FEDER Una manera de hacer Europa funded by MCIN/AEI/10.13039/501100011033/ under Grant CONFIA PID2021-122916NB-I00. The work of Yamisleydi Salgueiro was supported in part by the Department of Industrial Engineering, Faculty of Engineering, Universidad de Talca, Campus Curico, Chile, in part by CENIA under Grant FB210017, in part by Basal ANID, and in part by the super-computing infrastructure of the NLHPC under Grant ECM-02. The work of Agnieszka Jastrzebska was supported by the Warsaw University of Technology within the Excellence Initiative: Research University (IDUB) Program. Isel Grau and Gonzalo Na poles contributed equally to thework

Funders	Funder number
European Regional Development Fund	CONFIA PID2021-122916NB-I00
Universidad de Talca

Access to Document

10.1109/MCI.2023.3327834

MCI3327834Final author version , 4.29 MB
pdfFinal published version, 3.29 MBLicence: TAVERNE

Cite this

@article{5c4bb53e5b3d4be680b5f9b8c9111130,

title = "Forward Composition Propagation for Explainable Neural Reasoning",

abstract = "This paper proposes an algorithm called Forward Composition Propagation (FCP) to explain the predictions of feed-forward neural networks operating on structured classification problems. In the proposed FCP algorithm, each neuron is described by a composition vector indicating the role of each problem feature in that neuron. Composition vectors are initialized using a given input instance and subsequently propagated through the whole network until reaching the output layer. The sign of each composition value indicates whether the corresponding feature excites or inhibits the neuron, while the absolute value quantifies its impact. The FCP algorithm is executed on a post-hoc basis, i.e., once the learning process is completed. Aiming to illustrate the FCP algorithm, this paper develops a case study concerning bias detection in a fairness problem in which the ground truth is known. The simulation results show that the composition values closely align with the expected behavior of protected features. The source code and supplementary material for this paper are available at https://github.com/igraugar/fcp.",

author = "Isel Grau and Gonzalo N{\'a}poles and Marilyn Bello-Garc{\'i}a and Yamisleydi Salgueiro and Agnieszka Jastrz{\c e}bska",

year = "2024",

month = feb,

doi = "10.1109/MCI.2023.3327834",

language = "English",

volume = "19",

pages = "26--35",

journal = "IEEE Computational Intelligence Magazine",

issn = "1556-603X",

publisher = "Institute of Electrical and Electronics Engineers",

number = "1",

}

TY - JOUR

T1 - Forward Composition Propagation for Explainable Neural Reasoning

AU - Grau, Isel

AU - Nápoles, Gonzalo

AU - Bello-García, Marilyn

AU - Salgueiro, Yamisleydi

AU - Jastrzębska, Agnieszka

PY - 2024/2

Y1 - 2024/2

N2 - This paper proposes an algorithm called Forward Composition Propagation (FCP) to explain the predictions of feed-forward neural networks operating on structured classification problems. In the proposed FCP algorithm, each neuron is described by a composition vector indicating the role of each problem feature in that neuron. Composition vectors are initialized using a given input instance and subsequently propagated through the whole network until reaching the output layer. The sign of each composition value indicates whether the corresponding feature excites or inhibits the neuron, while the absolute value quantifies its impact. The FCP algorithm is executed on a post-hoc basis, i.e., once the learning process is completed. Aiming to illustrate the FCP algorithm, this paper develops a case study concerning bias detection in a fairness problem in which the ground truth is known. The simulation results show that the composition values closely align with the expected behavior of protected features. The source code and supplementary material for this paper are available at https://github.com/igraugar/fcp.

AB - This paper proposes an algorithm called Forward Composition Propagation (FCP) to explain the predictions of feed-forward neural networks operating on structured classification problems. In the proposed FCP algorithm, each neuron is described by a composition vector indicating the role of each problem feature in that neuron. Composition vectors are initialized using a given input instance and subsequently propagated through the whole network until reaching the output layer. The sign of each composition value indicates whether the corresponding feature excites or inhibits the neuron, while the absolute value quantifies its impact. The FCP algorithm is executed on a post-hoc basis, i.e., once the learning process is completed. Aiming to illustrate the FCP algorithm, this paper develops a case study concerning bias detection in a fairness problem in which the ground truth is known. The simulation results show that the composition values closely align with the expected behavior of protected features. The source code and supplementary material for this paper are available at https://github.com/igraugar/fcp.

UR - https://www.scopus.com/pages/publications/85182885890

U2 - 10.1109/MCI.2023.3327834

DO - 10.1109/MCI.2023.3327834

M3 - Article

SN - 1556-603X

VL - 19

SP - 26

EP - 35

JO - IEEE Computational Intelligence Magazine

JF - IEEE Computational Intelligence Magazine

IS - 1

M1 - 10384511

ER -

Forward Composition Propagation for Explainable Neural Reasoning

Abstract

Funding

Access to Document

Other files and links

Fingerprint

Code: Forward Composition Propagation

Forward Composition Propagation for Explainable Neural Reasoning

Cite this