The rise of Deep Learning in the past two decades has prompted research into solutions to help improve Alzheimer's diagnosis based on neuroimaging data. As such, a wide variety of different techniques have been used, but a clear turn towards the use of Convolutional Neural Networks (CNN) has been observed in the last decade. To effectively predicate Alzheimer's Disease (AD), this paper proposed a two stage method. The first stage involves learning the best representation of the training data using an improved sparse autoencoder (SAE), an unsupervised neural network. The second stage involves using a 3D-Convolutional Neural Network (3D-CNN) to differentiate between the health status and diseased status based on the learned records and MRI scan of the brain. The SAE was optimized so as to train an efficient model. We report on experiments using the ADNI data set involving 897 historical scans. We demonstrate that using 3D convolutional neural networks with sparse auto encoder outperform several other classifiers stated in the literature.