Full text
| |
| Author(s): |
Carlos da Silva dos Santos
Total Authors: 1
|
| Document type: | Doctoral Thesis |
| Press: | São Paulo. |
| Institution: | Universidade de São Paulo (USP). Instituto de Matemática e Estatística (IME/SBI) |
| Defense date: | 2010-06-25 |
| Examining board members: |
Roberto Hirata Junior;
Junior Barrera;
Roberto Marcondes Cesar Junior;
Joao Eduardo Kogler Junior;
Nelson Delfino D\'Ávila Mascarenhas
|
| Advisor: | Roberto Hirata Junior |
| Abstract | |
This work addresses the task of feature extraction for classification when both labels and features are binary. Our approach aims to build features with reduced interaction effects, thus relieving the classifiers from dealing explicitly with interactions. We introduce a new technique for building binary features by means of a parity calculating matrix that changes the coordinates of a binary vector. That matrix can be built purposefully for manipulating specific information theoretic measures. The resulting transform gives rise to groups of binary variables. A new algorithm for independent component analysis of binary features is proposed, based on this technique. In the context of classification, a new algorithm is presented that reduces the conditional dependence of features, given the label. A third algorithm presented in this text reduces the Interaction Information between pairs of features, a measure associated with redundancy or collaboration among predictors. This algorithm is used in the context of designing two-level operators for binary images. This type of operator combines the responses of several first-level operators to create the output image. In this case, the interaction reducing algorithm handles the division of the image into subregions. A complete framework for designing these operators is provided, involving feature selection and model comparison techniques. Experimental results show that the proposed method generates more accurate operators, compared to the single-level ones. (AU) | |
| FAPESP's process: | 05/04614-7 - Classification techniques in cardiac diagnosis based on motion quantification of nuclear medicine images |
| Grantee: | Carlos da Silva dos Santos |
| Support Opportunities: | Scholarships in Brazil - Doctorate |