Age and gender classification on human face images has been a topic with great interest by the research community and is used in a variety of real-world applications recenlty. Although several machine learning and deep learning models offered to solve the given task have achieved impressive results on benchmark datasets, there is still lacking performance on real world cases due to the large intra-class variety of face images (lighting variation, low-resolution, posing, scale, expressions, occlusions). To address the mentioned problems, we propose a new framework - self-normalized multimodel deep learning model which obtained high accuracy rate in performance by learning representations. To further increase representation and classification capability of the proposed framework, We benefited from the self-normalizing layers to the input and suitable parameter initialization methods. The model is trained on the Adience benchmark for age and gender prediction to label input images into eight class ranges of age and two classes of gender. With the proposed framework following by pre-processing and self-normalizing steps, experimental results demonstrate improved accuracy in both age and gender prediction, reaching state of the art performance in gender estimation.