The project addresses the problem of automatized classification of spectroscopic data by artificial neural networks. The biggest challenges in this field are a huge over-parametrization (and computational cost), interpretability and model generalizability. The proposed solution is based on a novel strategy for weight initialization of the network, using spectra modeling. This results in a rapid reduction of parameters and training time, with slightly better performance and interpretability.

CEITEC-J-21-7542

Czech

Vrábel Jakub, Ing. - principal person responsible
Kaiser Jozef, prof. Ing., Ph.D. - fellow researcher

Advanced instrumentation and methods for material characterization
- responsible department (5.3.2021 - not assigned)
Central European Institute of Technology BUT
- responsible department (14.2.2021 - 5.3.2021)
Advanced instrumentation and methods for material characterization
- internal (1.1.2021 - 31.12.2021)
Central European Institute of Technology BUT
- beneficiary (1.1.2021 - 31.12.2021)

VRÁBEL, J.; KÉPEŠ, E.; POŘÍZKA, P.; KAISER, J. Artificial neural network weights penalization and initialization for spectroscopic data. 2021.
Detail

Responsibility: Vrábel Jakub, Ing.