Publication detail

Towards interpretability of ANNs for spectroscopic data: inductive bias, lottery tickets, and input optimization

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

Original Title

Towards interpretability of ANNs for spectroscopic data: inductive bias, lottery tickets, and input optimization

Type

abstract

Language

English

Original Abstract

The interpretability of Artificial Neural Network (ANN) –based models remains a challenging task not only for spectroscopic data. We study and compare several distinct approaches that provide an improved understanding of the model’s (fully-connected network) predictions in supervised tasks and relate them to spectroscopic expertise. Namely, a weight initialization by modeled spectra and custom loss function penalization enable interpretation of the first hidden layer of the network. Additionally, lottery tickets (i.e. iteratively pruned networks) are used to reveal a local structure and positions of relevant features. The results are critically evaluated and compared to a baseline approach (feature visualization by input optimization).

Keywords

spectroscopic data, machine learning, interpretability, artificial neural networks

Authors

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

Released

1. 8. 2022

BibTex

@misc{BUT180067,
  author="Jakub {Vrábel} and Erik {Képeš} and Pavel {Pořízka} and Jozef {Kaiser}",
  title="Towards interpretability of ANNs for spectroscopic data: inductive bias, lottery tickets, and input optimization",
  year="2022",
  note="abstract"
}