Probabilistic Shrinkage Priors for Calibrated Representation Spaces
International Journal of Computational and Biological Sciences · 26.09.2026 · 3 (4) · 96-106
DOI: 10.66238/ijcbs125 ↗Kurz gefasst
Fachartikel in der Zeitschrift International Journal of Computational and Biological Sciences, online erschienen am 26.09.2026. Verfasst von Elena Khan. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-ND 4.0).
Abstract (Originalsprache)
Modern deep learning architectures excel at learning complex representations from high-dimensional data, achieving remarkable predictive performance across various domains. However, these models frequently suffer from poor uncertainty calibration, producing overconfident predictions, particularly when confronted with out-of-distribution samples or noisy inputs. This lack of calibration poses significant challenges in safety-critical applications such as medical diagnostics and autonomous navigation. To address this fundamental limitation, this paper introduces a novel framework utilizing probabilistic shrinkage priors to enforce regularization directly within the latent representation space. By imposing heavy-tailed hierarchical priors, such as the horseshoe distribution, on the feature embeddings, our method induces sparsity while preserving highly informative latent dimensions. This selective shrinkage prevents the network from memorizing spurious correlations and bounds the Lipschitz constant of the mapping, naturally leading to a more calibrated uncertainty surface. We develop a variational inference scheme to optimize the intractable posterior distribution, effectively balancing representation capacity with regularization. Through extensive theoretical analysis and empirical evaluation, we demonstrate that the proposed probabilistic shrinkage approach substantially reduces the Expected Calibration Error and improves predictive log-likelihoods compared to standard Gaussian priors and deterministic regularization techniques. The results establish that structured probabilistic priors are a critical mechanism for developing trustworthy deep learning systems capable of robust and calibrated uncertainty quantification.
Zitieren
APA
Khan, E. (2026). Probabilistic Shrinkage Priors for Calibrated Representation Spaces. International Journal of Computational and Biological Sciences, 3(4), 96-106. https://doi.org/10.66238/ijcbs125
Vancouver
Khan E. Probabilistic Shrinkage Priors for Calibrated Representation Spaces. International Journal of Computational and Biological Sciences. 2026;3(4):96-106. doi:10.66238/ijcbs125
BibTeX
@article{Khan2026,
author = {Khan, Elena},
title = {Probabilistic Shrinkage Priors for Calibrated Representation Spaces},
journal = {International Journal of Computational and Biological Sciences},
year = {2026},
volume = {3},
number = {4},
pages = {96--106},
doi = {10.66238/ijcbs125}
}
EUVIMED