Publication detail

Artificial neural network weights penalization and initialization for spectroscopic data

VRÁBEL, J. KÉPEŠ, E. POŘÍZKA, P. KAISER, J.

Original Title

Artificial neural network weights penalization and initialization for spectroscopic data

Type

abstract

Language

English

Original Abstract

Nowadays, Artificial Neural Networks (ANNs) are among the most utilized techniques for the advanced processing of spectroscopic data. However, several problems have emerged from such a broad adoption that are limiting their performance and trustworthiness. The most significant shortcomings are: 1) “blackbox-like” utilization of ANNs, 2) overtrained and overparametrized models, and 3) a direct transition of ANN architecture from different tasks and data types (e.g. image processing). In this work, we mainly focus on the third mentioned problem and propose several adjustments to the architecture and learning process of ANNs, which are suitable for spectroscopic data. Our approach is based on the unique properties of spectroscopic data and their direct exploitation in form of special weight initialization strategies or penalizations of the loss function. Adjusted models provide improved analytical performance and interpretability.

Keywords

spectroscopic data, artificial neural networks, machine learning, weights, initialization

Authors

VRÁBEL, J.; KÉPEŠ, E.; POŘÍZKA, P.; KAISER, J.

Released

30. 11. 2021

BibTex

@misc{BUT175292,
  author="Jakub {Vrábel} and Erik {Képeš} and Pavel {Pořízka} and Jozef {Kaiser}",
  title="Artificial neural network weights penalization and initialization for spectroscopic data",
  year="2021",
  note="abstract"
}