Avançada análise do uso de novos vetores-alvo em MLPs de alta performance

Abstract

This work proposes an advanced analysis for MLP performance improvement by adopting new target vectors. Firstly, a mathematical study is done to evaluate the influence of VBOs used as target vectors on MLP training. The VBOs provide the largest possible Euclidean distance between them to improve the training and generalization capability of MLPs. The largest distance inducement between points from output space leads to direct correspondence on pattern classification improvement. The various types of target vectors such as VBNs, VBCs and VNOs are adopted for training of MLP models and their performances are compared with the model trained by using VBOs. The mathematical evidences of performance improvement were found on weight updating refinement from backpropagation error stage of the algorithm. This particular refinement for training with VBOs is useful to preserve the features of each pattern due to noise interference reduction during the training process from a pattern to another. Following the mathematical study, more advanced experimental analysis using VBOs with two databases for pattern recognition is performed. The first database is related to the handwritten digits for comparing the performances of MLPs trained by adopting VBCs and VNOs with the performance of MLP trained by adopting VBOs. The results showed higher classification rates for the MLP trained with VBOs. The second database is constituted by human iris images in order to perform the comparison of MLP performances using conventional target vectors and new target ones represented by VBOs. Besides the high performance of MLPs trained with VBOs on recognition rates, it was concluded that the use of new target vectors provides high recognition rates with low tolerance for epoch trainings leading to the consequent low computational load for pattern processing.