Artificial Neural Networks (NNs) possess a great potential for automatic classification: they can be trained to classify successfully noisy, incomplete, or inconsistent data. However, an automation degree of existing NNs is relatively low because of several limitations. Usually, NNs require manual configuration: a large number of tuning parameters and the need for choosing an appropriate network topology lead to lengthy experimental cycles performed to find the best settings for a certain classification task. Additionally, many NNs need retraining in respect to changing input data and therefore cannot learn automatically in non-stationary environments. Another disadvantage of NNs is their «black-box» nature: they are not suitable for verification or interpretation by humans. A low degree of automation prevents employment of NNs in many practical applications when the users cannot be involved in the training and classification processes. Therefore the initial goal of the presented research was to develop a neural classifier that would be well-suited for automation.FĂŒr die Verwendung von kĂŒnstlichen neuronalen Netzen als automatischen Klassifikatoren sprechen ihre LernfĂ€higkeit, Modellfreiheit und NichtlinearitĂ€t der Datenapproximation. Ein zusĂ€tzlicher Vorteil ist ihre FĂ€higkeit zur Verarbeitung unvollstĂ€ndiger, verrauschter oder inkonsistenter Information. Trotzdem beeintrĂ€chtigen drei wichtige Faktoren die Automatisierbarkeit neuronaler Algorithmen: die Konfiguration eines Netzes benötigt oft entweder viel Fachwissen und Erfahrung oder eine Phase zeitaufwĂ€ndiger Parameteroptimierung; die neuronale Klassifikatoren zeigen ein sogenanntes «Black-Box»-Verhalten, d.h. das angelernte Wissen ist nicht interpretierbar. AuĂerdem sind nur wenige Algorithmen so stabil, dass sie in der Lage sind, in nicht stationĂ€ren Anwendungen zu lernen.

- 249 pages
- English
- PDF
- Available on iOS & Android
eBook - PDF
About this book
Trusted by 375,005 students
Access to over 1.5 million titles for a fair monthly price.
Study more efficiently using our study tools.
Information
Print ISBN
9783865370679
Edition
1