Deep Learning is a new branch of Machine Learning that focuses on applying Artificial Neural Networks to obtain more cluttered decision boundaries. However, the supervised Neural Networks are sensitive to presentation order, architecture configuration, and complex shapes. In addition, the learning instability causes large changes in its performance on training samples. In our study, we proposed a new model for RBF Neural Network tuning by using fuzzy consensus clustering and model selection. Experimental studies showed that our Deep Learning model, which is named DeepRBF, yielded better recognition accuracy than other supervised learning algorithms. In addition, our typical initialization scheme speeds up the learning convergence and avoids the local minimum.