Transfer Learning for Gender and Age Prediction

Abstract

In this work, we propose a transfer learning pipeline for gender and age prediction using images from IMDB-WIKI dataset. Firstly, we freeze all layers in pre-trained ImageNet models. Then, the models are trained for four stages with scheduled learning rates and the blocks of layers are unlocked consecutively in accordance to the schedule. We apply multi-output neural network paradigm to predict age and gender simultaneously and the final loss function is based on the combination of age and gender losses. In our approach, the model has better performance than that of the non-pre-trained model because the later stages of our models reuse features extracted from the pre-trained early stages.