Breast cancer (BC) is one of leading cause of cancer death among women. Early detection of this pathology allows to define, for patients, the appropriate treatment which improve the possibility of survival. Consequently, the earlier it is detected, the better the survival rate will be. In last years, big interest has been made on applying deep learning neural networks to BC analysis, detection and classification. Several interesting results have been found but they still needing more improvement and validation. In this frame is located our work which consists on developing and comparing different deep leaning approaches for earlier detection and classification of BC from mammogram images. Different new deep learning techniques are developed resulting on a considerable improvement in the rate of the classification accuracy with more robust models. Simulations are carried out on two data bases illustrating that the new proposed approaches provide significant evaluation metrics in comparison with several commonly known transfer learning architectures.