TY - GEN
T1 - On the Relationship Between Neural Gradients and Model Reasoning
AU - Richards, Jéssica
AU - Pedreira, Carlos Eduardo
AU - Wanner, Elizabeth Fialho
AU - Marcelino, Carolina Gil
N1 - Copyright © The Author(s), under exclusive license to Springer Nature Switzerland AG 2026. This version of the article has been accepted for publication, after peer review and is subject to Springer Nature’s AM terms of use [ https://www.springernature.com/gp/open-research/policies/accepted-manuscript-terms ] but is not the Version of Record and does not reflect post-acceptance improvements, or any corrections. The Version of Record is available online at: https://doi.org/10.1007/978-3-032-25552-5_8
PY - 2026/7/2
Y1 - 2026/7/2
N2 - As machine learning increasingly shapes decisions that directly impact people’s lives, the demand for transparency and accountability has grown. Explainable Artificial Intelligence (XAI) has emerged as a field dedicated to making the reasoning of autonomous decision systems clearer and more interpretable. Among the many approaches to XAI, gradient-based explanations use a model’s internal gradients to highlight feature importance. In this work, we investigate whether a model’s gradients reveal consistent patterns that reflect its reasoning process. To keep the problem tractable, given the exponential growth in neural network dimensionality and corresponding gradients, we conducted our experiments using a Multilayer Perceptron as a controlled, toy example. For statistical analysis, we applied the Mantel and Energy tests. Our results indicate that gradient patterns emerge as intrinsic properties of the model’s architecture and the dataset on which it is trained.
AB - As machine learning increasingly shapes decisions that directly impact people’s lives, the demand for transparency and accountability has grown. Explainable Artificial Intelligence (XAI) has emerged as a field dedicated to making the reasoning of autonomous decision systems clearer and more interpretable. Among the many approaches to XAI, gradient-based explanations use a model’s internal gradients to highlight feature importance. In this work, we investigate whether a model’s gradients reveal consistent patterns that reflect its reasoning process. To keep the problem tractable, given the exponential growth in neural network dimensionality and corresponding gradients, we conducted our experiments using a Multilayer Perceptron as a controlled, toy example. For statistical analysis, we applied the Mantel and Energy tests. Our results indicate that gradient patterns emerge as intrinsic properties of the model’s architecture and the dataset on which it is trained.
KW - Explainable Artificial Intelligence
KW - Gradients
KW - Multilayer Perceptron
UR - https://link.springer.com/chapter/10.1007/978-3-032-25552-5_8
UR - https://www.scopus.com/pages/publications/105047681339
U2 - 10.1007/978-3-032-25552-5_8
DO - 10.1007/978-3-032-25552-5_8
M3 - Conference publication
SN - 9783032255518
VL - 16421
T3 - Lecture Notes in Computer Science (LNCS)
SP - 119
EP - 126
BT - From Data to Models and Back: 13th International Symposium, DataMod 2025, Toledo, Spain, November 10–11, 2025, Revised Selected Papers
A2 - Lestingi, Livia
A2 - Salaün, Gwen
A2 - Requeno Jarabo, José Ignacio
PB - Springer, Cham
T2 - 13th International Symposium on From Data Models and Back, DataMod 2025
Y2 - 10 November 2025 through 11 November 2025
ER -