A new classifier topology – generalized classifier neural network and its hardware implementations

In this study, a new radial basis function based classification neural network named as generalized classifier neural network is developed and implemented on field programmable gate arrays. Moreover, introduced neural network is improved by firstly, adapting logarithmic learning method (logarithmic learning for generalized classifier neural network) and then, utilizing a new smoothing parameter calculation method (one pass learning for generalized classified neural network) proposed in this study. Since one pass learning eliminates the training step of generalized classifier neural network, it increases the hardware implementability of generalized classifier neural network; hence, it is implemented on field programmable gate array with off-board learning. Introduced classifier and learning methods are tested on 15 UCI machine learning repository datasets as 10-fold cross validation and compared with frequently used neural network methods in the literature. Test results proved the efficiency of classifier and learning methods. Generalized classifier neural network evolved to be used in real time applications by exploiting the one pass learning method. In addition to software tests, proposed hardware is simulated and implemented on Xilinx ML605 Evaluation board as 10-fold cross validation testing for 8 UCI machine learning repository datasets. Simulation and implementation results show that generalized classifier neural network hardware is working without accuracy loss. On the other hand, due to the resource limit of hardware, training data are divided into sets and applied to the hardware set by set. Hence, parallelism of hardware is limited to the number of training data in the set and speed gain is lower than expected.