Training of mixed deep classifiers of symbolical rules and neural networks - Training von gemischten Regel- und neuronalen Netzwerk-Klassifizierern
| Type of publication: | Mastersthesis |
| Citation: | ma:straninger |
| Type: | Master thesis |
| Year: | 2016 |
| Month: | December |
| School: | TU Darmstadt, Knowledge Engineering Group |
| URL: | /lehre/arbeiten/master/2016/Straninger_Andreas.pdf |
| Abstract: | This thesis deals with combining a decision list classifiers with neural networks in order to gain a mixed classification model, which can be further transformed into a pure multi-layered decision list classifier. The resulting classifier has the potential of increase the accuracy especially for deep network topologies. In contrast to neural networks, the rules in decision lists are better understandable by humans. This thesis shows the relations of the mixed network to ensemble trainers (especially gradient boosting), fuzzy neural networks and other descriptive rule learning algorithms. The contribution of this thesis is a method for performing an online training of neural networks and decision lists. For this purpose, a way of training decision lists with small portions of data (mini-batches) instead of complete data sets is introduced. Further contributions include the way of combining rule learning algorithms with neural networks similar to gradient boosting and a binary backpropagation scheme for symbolical methods. During evaluation with the MNIST dataset, it is shown that existing classifiers like C4.5 and RIPPER benefit from the results of the decision list layers and can increase their performance. In comparison to existing approaches, a model with a reduced complexity can be obtained. |
| Userfields: | betreuer={ELM} |
| Keywords: | |
| Authors | |
|
Attachments
|
|
|
|
|
Topics
|
|
|
|
