Although deep learning architectures are nowadays used in several research fields where automatized investigation of large scale datasets is required, the intrinsic mechanisms of deep learning networks are not fully understood yet. In this paper, a new approach for characterizing how information is processed within convolutional neural networks (CNNs) is introduced. Taking advantage of an analysis based on information theory, we are able to derive an index that is associated with the degree of maximum information extraction a CNN can obtain under ideal circumstances as a function of its hyperparameters setup and of the data to be explored. Experimental results on remote sensing datasets show the robustness of our approach. The outcomes of our analysis can be used to optimize the design of CNNs and maximize the information that can be obtained for the considered problem.

On the Optimal Design of Convolutional Neural Networks for Earth Observation Data Analysis by Maximization of Information Extraction

Iannelli G. C.;Gamba P.
2020-01-01

Abstract

Although deep learning architectures are nowadays used in several research fields where automatized investigation of large scale datasets is required, the intrinsic mechanisms of deep learning networks are not fully understood yet. In this paper, a new approach for characterizing how information is processed within convolutional neural networks (CNNs) is introduced. Taking advantage of an analysis based on information theory, we are able to derive an index that is associated with the degree of maximum information extraction a CNN can obtain under ideal circumstances as a function of its hyperparameters setup and of the data to be explored. Experimental results on remote sensing datasets show the robustness of our approach. The outcomes of our analysis can be used to optimize the design of CNNs and maximize the information that can be obtained for the considered problem.
2020
978-1-7281-6374-1
File in questo prodotto:
Non ci sono file associati a questo prodotto.

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/11571/1439704
Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus 0
  • ???jsp.display-item.citation.isi??? 0
social impact