Decomposing A Neural Learner
Abstract
Neural learners for classification and regression are often criticized as black boxes. We propose a probabilistic decomposition of neural learners that accounts for heterogeneous predictability across the feature space. A training procedure combining the EM algorithm, stochastic gradient descent, and maximum likelihood estimation is developed to jointly learn the predictable component and an input-dependent reliability score. We show that the optimal decomposed learner recovers the response distribution within the predictable component, while the reliability score characterizes the predictability of individual observations. We further extend the framework to incorporate localized feature transformations. Experiments on synthetic and real-world datasets demonstrate the effectiveness of the proposed framework.
Related
- AI-Integrated Computational Framework for Pronunciation Error Analysis Through Native–Learner Speech Comparison
- Graph-Enhanced Knowledge Tracing for Explainable Learner Modeling: A Systematic Literature Review
- A Borel Concept Class of VC Dimension One with a Non-PAC Consistent Learner in ZFC
- Decomposing the Action of Macdonald Operators for Classical Types
- Decomposing multigraphs into two Class I subgraphs with prescribed maximum degrees
BibTeX
@misc{bi2026decomposing,
title = {{Decomposing A Neural Learner}},
author = {Zelong Bi},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.14388},
eprint = {2610.14388},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.14388},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Decomposing A Neural Learner
AU - Zelong Bi
PY - 2026
DA - 2026/10/09
PB - arXiv
DO - 10.48550/arxiv.2610.14388
UR - https://arxiv.org/abs/2610.14388
AB - Neural learners for classification and regression are often criticized as black boxes. We propose a probabilistic decomposition of neural learners that accounts for heterogeneous predictability across the feature space. A training procedure combining the EM algorithm, stochastic gradient descent, and maximum likelihood estimation is developed to jointly learn the predictable component and an input-dependent reliability score. We show that the optimal decomposed learner recovers the response distribution within the predictable component, while the reliability score characterizes the predictability of individual observations. We further extend the framework to incorporate localized feature transformations. Experiments on synthetic and real-world datasets demonstrate the effectiveness of the proposed framework.
ER -