Predicting software defect is considered as one of the most dynamic research field in the domain of software engineering. Prediction of defects can improve the software quality by indicating the particular areas in advance where faults are more likely to occur. Software defect prediction models have accomplished the significant acceptance in software industry in the last decade. Prediction and prevention of defects in beginning stages of software?s development can reduce the whole development time and cost by reducing the testing efforts. The widely used data mining models for defect prediction includes one or more of the following machine learning techniques: ?Logistic Regression (LR), Naive Bayes (NB), K-Nearest Neighbor (k-NN), Neural Network (NN), Decision Tree (DT), Support Vector Machine (SVM), and Random Forest (RF)?. According to latest research, Neural Networks come up with higher accuracy than other mathematical and statistical models. Artificial Neural Network is widely accepted supervised learning approach to deal with the prediction problems in multiple domains of the software engineering such as ?effort estimation, cost estimation and software defect prediction?. Many researchers have proposed models based on Neural Networks for software defect prediction, however wide comparison with benchmark datasets were lacked. Moreover, in most of researches, WEKA was used which has limited parameters for developing Neural Networks. In this study, a model will be proposed by using ANNs for the effective defects prediction with Feed-Forward and Feed-Back techniques in MATLAB, which provides more parameters for the development and tuning the Neural Networks. An empirical comparison of ?Levenberg-Marquardt, Bayesian Regularization, Scaled Conjugate Gradient and BFGS Quasi-Newton? back propagation training algorithms will be performed. The performance of proposed model will be analyzed by using benchmark datasets available in NASA MDP repository. A GUI based Neural Network Simulator will developed to assess and tune the performance of our ANNs models which will provide the options to select appropriate learning function along with different combinations of hidden layers and the numbers of neurons in each layer.