Static Field Approach for Pattern Classification
Abstract:
Abstract. Recent findings in pattern recognition domain show that dramatic improvement of the recognition rate can be obtained by application of the fusion systems utilizing many different and diverse classifiers for the same task. Apart from a good individual performance of individual classifiers the most important factor is the useful diversity they exhibit. In this work we present an example of a novel non-parametric classifier design, which shows substantial level of diversity with other commonly used classifiers. In our approach inspiration for the new classification method has been found in the physical world. Namely we considered training data as particles in the input space and exploited the concept of a static field acting upon the samples. Specifically every single data point used for training was a source of a central field curving the geometry of the input space. The classification process is presented as a translocation in the input space along the local gradient of the field potential generated by the training data. The label of a training sample to which it converged during the translocation determines the eventual class label of the new data point. Based on selected simple fields found in nature, we show extensive examples and visual interpretations of the presented classification method. The practical applicability of the new model is examined and tested using well-known real and artificial datasets. 1
Citations
| 932 | Pattern Classification – Duda, Hart, et al. - 2000 |
| 42 | Combining Artificial Neural Nets: Ensemble and Modular Multi-net Systems – Sharkey - 1999 |
| 37 | Mutual information in learning feature transformations – Torkkola, Campbell - 2000 |
| 21 | Combining diverse neural nets – Sharkey, Sharkey - 1997 |
| 15 | Ten measures of diversity in classifier ensembles: limits for two classifiers – Kuncheva, Whitaker - 2001 |
| 13 | Complexity, Entropy and the Physics of Information – Zurek - 1990 |
| 12 | Analysis of the correlation between majority voting error and the diversity measures in multiple classifier systems – Ruta, Gabrys - 2001 |
| 5 | An electric approach to independent component analysis – Hochreiter, Mozer - 2000 |
| 3 | Folger T.A.: Fuzzy Sets – Klir - 1988 |
| 2 | III, Xu D.: Information Theoretic Learning – Principe, Fisher - 2000 |
